HIGH SPEED DATA CLASSIFICATION SYSTEM
An optical network packet classification architecture is disclosed that addresses the packet classification requirements for OC-768 optical routers and beyond. The herein disclosed system is used for ultra-high speed packet classification of optical data at either the serial data stream level for maximum performance, or after it has been converted into parallel words of data. The presently preferred embodiment of the invention provides a system that operates in the receive path, where electronic data are provided by the optical interface to the data framer. The invention incorporates unique features into a traditional optical data framer chip and relies on a complex ASIC to permit the user to differentiate between up to 10,000 different patterns at ultra-high speeds. One purpose of the general purpose system disclosed herein is to eliminate the need for costly and power consumptive content addressable memory systems, or customer pattern specific ASICs, to perform network packet classification. The system operates on a principle of adaptive programmable randomization to permit a differentiation between the input vectors to be made. The invention dramatically reduces the processing burden required by high-speed optical routers or switches.
This application is a continuation of U.S. patent application Ser. No. 09/944,572, filed Aug. 30, 2001, which is incorporated herein in its entirely by this reference thereto.
BACKGROUND OF THE INVENTION1. Technical Field
The invention relates to computer networks. More particularly, the invention relates to an information processing system.
2. Description of the Prior Art
Communication between computers over the Internet can be compared to the delivery of mail and packages by the United States Postal Service. Users access the Internet through a variety of options, e.g. phone modems, DSL modems, cable modems, T-1 lines, local area networks, wireless networks, and wide area networks.
In the world of the U.S. Postal Service, access to the mail system could be through a mail-slot in the door of your home, a mailbox at the street in front of your home, a post office box on a street corner, a post office counter, or a post office box. By analogy, each user of the Internet is assigned an address, and the Internet infrastructure learns how to deliver messages intended for them.
In the world of the Postal Service, the zip code, city, street, and street number are used progressively to determine how to route and deliver the mail. Users of the Internet rely on various networking protocols to transfer messages between computers.
In the world of the Postal Service, the protocols for delivering the mail include First Class Delivery, Next Day Air, Parcel Post, and Bulk. In the world of the Internet, messages are sent in packets, as opposed to the letters that are sent in the world of the Post Office. These packets contain information necessary for delivery, and this information is found in the Packet Header. This packet header includes the recipient's addresses and the sender's address, as well as the delivery method and style of message. The packet header is comparable to all of the information that is visible on the outside of a letter or package, i.e. recipient's address, return address, mail type, and specific handling instructions, such as FRAGILE. The remainder of an Internet packet contains user data. This user data is comparable to what is found inside an envelope or package. The Internet infrastructure has no more need to see the user data to route and deliver the message to the intended computer accurately than the post office has to open the mail it handles to figure out where to send it. Table A below shows a typical Internet packet.
As computers are tied together over the World Wide Web, the physical connections between the computers look like a giant spider web. The thick strands of this web transfer huge numbers of packets between big cities to move them along their way. This is comparable to the air or truck traffic carrying millions of letters between postal hubs. At each connection point on the World Wide Web, a sorting function must be performed to determine which direction a message should be sent. This sorting of packets is similar to the process where high-speed postal sorters scan letters to determine their addresses and figure out which direction to send them. Sorting of data packets is often referred to as packet classification
An optical router is a device that has many input/output (I/O) ports or connections. Each I/O port connects through an optical fiber to another optical router, optical switch, or optical adapter that can be located a long distance geographically from the first device. In simplistic terms, the purpose of an optical router is to receive data packets on each I/O port, to interpret the headers within the packet, and to route the packet out the appropriate I/O port towards the destination computer. If an optical router is unable to sort packets quickly enough, packets backup and are potentially lost by the router. In such case, the Internet slows down and computer users may lose their connections. As more and more people use the Internet, the situation internal to the optical routers that makeup part of the Internet infrastructure can start to look as chaotic as the Post Office at Christmas time.
The goal of an optical router is to interpret the packet header for each received packet as fast as possible so that the packets can be sent out the correct I/O port. This avoids delays, backups, and potentially lost packets. One problem is that thousands of different users can be sending messages through a router at the same time, and the packets all need to be sorted and routed differently. Table B below shows how the number of possible headers that can be received increases dramatically as the number of bits in the packet header increases.
The problem of receiving a packet and identifying critical header information to decide where to route the packet is much like finding a needle in a haystack. Initially, routers used microprocessors and large lookup tables in memory to search for addresses and header information. Later, as data rates increased, system designers moved to content addressable memories (CAMs) to allow the received packet header to be compared to all previously analyzed packet headers simultaneously. The architecture of a CAM permits the user to apply the received header information to the memory and to determine to which location(s) it matches.
Because the performance of CAM's could not keep up with ultra-high speed router implementations, some manufacturers switched to custom ASICs (Aplication Specific Integrated Circuits) to evaluate packet headers in a rapid fashion.
Optical networking is a significant business opportunity because of the tremendous increases in data bandwidth requirements resulting from the increasing use of Internet. The capability of optical fibers to transmit and receive data exceeds the capability of electronic and electro-optical interface products to keep up with increasing data rates. Presently, OC-192 standard networks that operate at 10 Gbit/sec are beginning to be used. Presently available optical routers address the need attendant with processing and routing packets from OC-192 systems.
Existing Optical Network Packet Classification Schemes
High performance optical routers have been generally implemented using either CAMs or custom ASICS to perform packet classification. The custom ASIC approach must rely on filtering and interpreting some subset of possible packet data patterns to determine how to route packets. The approach is inflexible and may be difficult to scale with new standards and new protocols. The CAM approach is more flexible and is popular in high end routers. CAMs are designed to be cascaded so that greater numbers of data bits can be analyzed. CAMs are designed to permit various levels of “don't care” functionality that has increased their flexibility and usefulness.
CAM Based Classification Systems
A typical router is shown in
Custom ASIC Solution (Juniper Networks ASIC2)
Juniper Networks (Sunnyvale, Calif.) provides high performance routers that use a custom ASIC solution that is marketed as the Juniper Networks ASIC2. The Juniper Networks ASIC2 in conjunction with the Juniper “Junos” software allows up to 40M packets/sec to be forwarded in the Juniper system. From Juniper's data sheets, the following:
-
- Juniper's routers leave the packet in the shared memory and move only a packet pointer through the queues. When packets arrive they are immediately placed in distributed shared memory where they remain until being read out of memory for transmission. This shared memory is completely nonblocking, which in turn, prevents head-of-line blocking.
CAM and Custom ASIC Shortcomings for Packet Classification of OC-192 and Beyond
A variety of problems are beginning to plague CAM and customer ASIC based systems as data rates are moving to OC-192 (10 Gbits/sec) and OC-768 (40 Gbits/sec). Some of the biggest problems have to do with raw forwarding throughput, which is related to how many packets per second can be processed; latency, which is related to the absolute delay through a router; system power consumption; and board area. A key component of packet latency through a router is the time necessary to perform packet classification. As latency increases, the chances of experiencing upper level networking protocol timeouts for a packet increase.
Typical CAM structures have a width that is associated with how many bits the user desires to analyze, and a depth that is based on the number of possible patterns that the user wishes to differentiate between. CAMs are cascadable to meet both the width and depth that is required. The downside of cascading is that it costs money, increases board area, and increases power consumption. On the other hand, the ASIC2 solution from Juniper Networks does not appear to be cascadable. It appears to operate on data in the SRAM, and permits qualified searches on only certain fields and bits. This limits the ASIC2 solution approach when new search criteria are desired to be used.
Packet Classification Forwarding Rate and Latency Issues
The issues of forwarding rate and latency are intertwined and need to be addressed together. There are two significant architectural issues that affect forwarding rate and latency, i.e. the design of a packet's data flow through the system, and the underlying performance of the packet classification hardware.
In a CAM based system, such as that in
In addition to the delays and uncertainty associated with transferring the data from the data framer into the CAM memory, there is the delay of the CAM memory in processing the data once the final word has been presented. Typical CAM memories have delays of approximately 100 nsec from application of data to input match. This is expected to improve as CAM technologies improve, but is not likely to experience anything close to four times improvements as users move from OC-192 to OC-768. Due to this inherent access delay of CAM memories, the delay in receiving routing information becomes worse relative to data rate as speeds increase. This results in the need to increase queue's and storage depths to account for buffering data prior to knowing to where it should be routed.
Present CAM classification systems are claimed to operate at full line data rates. The problem is that they require packets to be received, staged, and then sent into the classification engine to determine an appropriate route or other required information. This delay increases the latency through the router for a packet to be sent. Eventually, this latency through a router can start to impact connections going through the router and can result in higher layer timeouts. As new CAM technologies are implemented, the focus is on increasing size and maintaining access time. Therefore, the access time is not scaling anywhere near as quickly as data rate.
In the case of the Juniper Network's ASIC2 solution (
Power Consumption and Board Area Issues.
The overall power consumption of a router system increases because the network processor speed must be increased to process higher data rates. In addition, CAM memories have a static current draw that must be accounted for and scaled up. As an example, a currently available network data base search engine using CAM technology draws 6 Amps @1.5 Volts running at 100 MHz. This is an extremely high 9 Watts on a single chip. This impacts the usability of this solution in applications where space is tight and power is limited. It is noted the increased power consumption also raises issues of heat dissipation that must be addressed.
As packet classification searches farther into a packet, such as to 512 or 1024 bits deep, CAM based solutions require multiple parts to be operated in parallel. This significantly increases power consumption and board area. In the case of the ASIC2 solution, increasing the depth of the classification requires an entirely new part to be developed. In addition, the ASIC2 solution could require greater memory bandwidths with higher speeds, which would entail more parts and larger ASICs.
It would be advantageous to provide an improved system for ultra-high speed packet classification of optical data that has been framed into a serial data stream.
SUMMARY OF THE INVENTIONThe herein disclosed invention provides a system that permits flexible, low latency, ultra-wide, and deep classification of high speed data. A presently preferred embodiment of the invention comprises an optical network packet classification architecture that addresses the packet classification requirements for OC-768 optical routers and beyond. Packet classification involves understanding the source and destination of a packet, as well as interpreting information within the packet header to determine what the optical network processor should do with the packet. As the data rates of optical networks move up to OC-768 and beyond, the job of performing packet classification is becoming increasingly more difficult. The approach used in the herein disclosed system allows for true “Light Speed” classification of optical data packets.
The herein disclosed system is used for ultra-high speed packet classification of optical data that has been framed into a serial data stream. The presently preferred embodiment of the invention provides a system that operates in the receive path, where electronic data are provided by the optical interface to the data framer. The preferred embodiment of the invention incorporates unique features into a traditional optical data framer chip and relies on a complex ASIC to permit the user to differentiate between up to 10,000 different patterns at light speed. One purpose of the general purpose system disclosed herein is to eliminate the need for costly and power consumptive content addressable memory systems, or customer pattern specific ASICs, to perform network packet classification. The system operates on a principle of adaptive programmable randomization to permit a differentiation between the input vectors to be made. The invention dramatically reduces the processing burden required by high-speed optical routers or switches.
The modified data framer that is used in the preferred system is referred to herein as the novel data framer: the complex ASIC that is used to control the adaptive programmable randomizer is referred to herein as the ASIC, both of which-are discussed in greater detail below.
BRIEF DESCRIPTION OF THE DRAWINGS
The herein disclosed system is used for ultra-high speed classification of data that have been organized into a serial or parallel data streams. The presently preferred embodiment of the invention provides a system that operates in the receive path, where electronic data are provided by the optical interface to a data framer. In one embodiment, the invention incorporates unique features into a traditional optical data framer chip and relies on a complex ASIC to permit the user to differentiate between up to 10,000 different patterns at light speed. One purpose of the general purpose system disclosed herein is to eliminate the need for costly and power consumptive content addressable memory systems, or customer pattern specific ASICs, to perform network packet classification. The system operates on a principle of adaptive programmable randomization to permit a differentiation between the input vectors to be made. The invention dramatically reduces the processing burden required by high-speed optical routers or switches.
The modified data framer that is used in the preferred system is also referred to as the novel data framer; the complex ASIC that is used to control the adaptive programmable randomizer is also referred to as the ASIC, both of which are discussed in greater detail below.
The herein disclosed optical network packet classification system analyzes the packet headers of messages sent over the Internet. At their lowest possible level, these packet headers are made up of a sequence of bits that are either a 1 or a 0. The address and networking information for each Internet packet are encoded into these header bits. Depending upon the method of sending the packet, it is possible for there to be many hundreds of bits in the packet header. These bits of data are transferred through optical fibers by pulsing light on or off.
For discussion purposes, the Table D below shows how a packet header would look if it were broken into discrete bits. In this example, the packet header is made up of only ten bits. The first bit to be sent using a light pulse is B0, and the last bit to be sent is B9. The sequence of light pulses corresponding to this packet header is 1110101001.
System Specifications
The following specifications (Table E) apply to a system comprising the presently preferred embodiment of the invention. The general specifications apply to the overall system performance. The bit masking specifications apply to the ability to mask, or ignore, bits in the received packet header. In the case of the herein disclosed system, there is extensive flexibility for masking bits in one operation and then removing the masking in a later step. The masking operation is comparable to looking at a piece of mail and first checking only the zip code to which it is being sent. The next step with the piece of mail might be to look at the city and the street address to determine which mail carrier should be given the letter.
Note 1
The initial system chipset is designed to handle 10,000 inputs. Architecturally, this could be increased to 40,000 inputs in a second version of the part. Ultimately, the architecture permits 80,000, 160,000, or more inputs with increased memory and ASIC sizes, but with no speed degradation. The programmable masking patterns of the system permit significantly more than 10,000 inputs to be handled effectively by the system. The 10,000 input number
Note 2
Expected values through the use of 7 nsec SRAMs in the system and based upon an average of 2.35 SRAM accesses per classification.
Note 3
Expected values through the use of 7 nsec SRAMs in the system and based upon a maximum of 6 SRAM accesses per classification.
The number of fixed block masking periods and the number of flexible masking bits can be increased with increased ASIC sizes. The fixed block masking is intended for a long sequence of bits that are always ignored, while the flexible masking bits are provided to deal with individual fields of bits.
Example of Fixed Block Masking
Table F shows a setup where the length of the packet header to be evaluated is 564 bits. This value is called the input length. All bits that are received after the 564th bit are not used in the packet classification because they exceed the Input Length. A fixed block masking period from bits 136-227 is used in the example. This means that bits falling within this bit window are not used in the input classification evaluation. As an example, the fixed block masking is comparable to ignoring the sender's address when trying to determine where mail should be sent.
Table G shows a setup where two, of a possible four, 32 bit flexible masking blocks are enabled. These blocks can be setup to fall on any valid 32 bit boundaries after the start of the header. In the example, the first flexible masking block #1 is setup to range from bits 96-127, which lies on a 32 bit boundary. The second flexible masking block #2 is setup to range from bits 352-383, which again lies on a 32-bit boundary. As an example, flexible masking is comparable to ignoring the destination city and street selectively for the recipient of a piece of mail in a rough check, and then looking at these only if the letter was sent using Next Day Air delivery.
Example of Flexible Masking (Sub-Block) Table H below shows bits 96-101 from flexible masking block #1 in Table G above. These five bits are shown to illustrate how the system permits bit level masking flexibility for any bit in a flexible masking block. To account for different possible masking configurations based upon the network protocols, addresses or fields that are received in the header, the presently preferred embodiment of the system provides eight selective masking patterns for each flexible masking block. The eight selective masking patterns are illustrated in Table H. In the case of selective masking pattern #1, bits 96, 98, 99, and 100 are masked (=1), while bits 97 and 101 are not masked (=0).
System Benefits
The system provides a range of performance, cost, power, and size benefits. The following are some of the key benefits provided by the herein disclosed system:
-
- Extremely Fast and Flexible Packet Classification—The system exceeds the speed of content addressable memory systems in packet processing. It permits extremely deep processing of bits in the packet header without impacting classification speed. When compared to custom solutions, the system provides deeper and more flexible processing at comparable throughputs.
- Unparalleled Search Latency—The system starts classifying a packet immediately after the last bit has been received. There is no overhead in performing data transfers to memory or to custom ASICs. This opens up new optical routing and switch architectures for ultra high performance.
- Low System Power Consumption—After the initialization process, the system has much lower power consumption than CAM or ASIC alternatives.
- Flexible and Programmable Masking—The system provides both fixed block and flexible masking. The flexible masking can be pre-programmed to go through sequential operations without external intervention. This feature is a significant advantage vis-à-vis content addressable memory approaches.
System Theoretical Background
The system uses a technique of adaptive, programmable, predictive, and sequential randomization to permit extremely rapid differentiation between a limited number of serial data bits. Optical networking relies on high speed transmission of digital data packets in a serial format. Optical routers and switches require that these serial packets be analyzed to determine the appropriate source and destination of the data packet so that they can be properly forwarded and at the correct priority level. All known present systems of packet analysis require serial data packets to be translated into a parallel format and then analyzed through the help of a combination of network processors, content addressable memories, custom ASICs, and high speed memories.
The randomization in the ASIC portion of the herein disclosed system is performed using compact, programmable feedback shift registers that are driven by the serial data stream. A general description as to how these programmable feedback shift registers are used in the system is provided below. The final state of these shift registers is used as an index into a memory array to determine which if any input data pattern has been matched. These shift registers require a simple register with exclusive OR feedback taps that can be programmed to be enabled or disabled. They have been designed to minimize power consumption, and the feedback tap enabling or disabling does not have an affect on the propagation delay of the feedback mechanism. The feedback mechanism has been kept simple, and minimal in terms of gate delays, to permit operation at extremely high serial data rates. More importantly, the various programmable feedback paths that are possible in the herein disclosed architecture have been selected specifically to guarantee that output values from one feedback value are uncorrelated to output values from another feedback value. This uncorrelated feature permits general probability theory to be used to evaluate the randomization of the data.
The predictive nature of the randomization comes from the fact that the randomization is pre-calculated for each possible input data pattern. This pre-calculation is done at the time that a new input data pattern is entered into the system for use. A critical feature of the system is that it implements a full hardware calculation of expected randomization outputs for each input that is applied. This hardware implementation allows many randomizer feedback values to be evaluated in real-time when a new input is applied. These randomizer output values are stored in memory for each randomizer feedback that is being considered at the time.
The adaptive randomization results from the system adjusting the randomization, over time, to handle the changing input data patterns that are to be analyzed in the best fashion. The high speed predictive nature of the system permits a significant number of possible randomization feedback paths to be maintained in memory at any time. The system can adjust the possible randomization feedback value after any packet has been received. This is done if the existing feedback randomization is significantly less ideal than another feedback randomization that has been evaluated. The system maintains statistics on all presently evaluated feedback randomization to determine the best randomization, as well as any randomization that may be no longer usable. When a randomization is no longer usable, the system can quickly bring all of the input data patterns back to evaluate other possible randomization patterns.
The sequential randomization that is permitted in the system results from the ability for the user to implement sequential masking operations on the input data. The system permits fixed or programmable masking of selected bit patterns within an input serial data stream. The masking operations of the system permit the user to pre-program a series of masking decisions that can result in a final input data pattern match.
Theoretical Randomization Probabilities
Detailed probability analysis is critical to an understanding of the system. The success of having a usable feedback randomization pattern, for a random set of inputs, depends upon the effective mapping of the input data patterns to output vectors by the randomizer. For practical implementations, with significant numbers of input vectors and reasonable sized memories, a system must be able to handle a limited number of cases where two or more input data patterns are mapped to the same output value. In the presently preferred embodiment of the system this is handled through a variety of methods including permitting a set value of multiple output cases where two, three, or four input data patterns map to the same output pattern. In addition, a secondary randomizer is used to separate between the multiple outputs so that the appropriate input can be determined.
The detailed theory behind evaluating any given randomizer pattern is presented below. This theory has been done in terms of the number of possible output states that are generated by the randomizer, and the number of possible input vectors that are being differentiated. The length of the input data patterns affects the predictive evaluation of the randomization outputs in hardware by the system, but it does not have a first order affect on the randomization probabilities. One part of the discussion below develops the theory to show the odds that a randomizer produces a specific case of a certain number of non-paired outputs, paired outputs, tripled outputs, or quadrupled outputs for a certain number of input data patterns. Another part of the discussion below evaluates how permitting various numbers of multiple outputs affects the possibility that a certain randomizer feedback is usable. For purpose of this discussion, unusable randomizer feedback occur when-too many input data patterns map to the same output, or when as a group, there are too many sets of input data patterns that map to different but common outputs. As an example, if 4000 input data patterns mapped to 2000 different outputs where there were two input data patterns for each output, and the system permitted only 1000 multiple outputs, the randomizer feedback is unusable.
Primary Randomizer feedback Selection Probabilities
The primary randomizer in the system is used to perform the mapping of each input data pattern to an output value. Given a number of input data patterns, there are always certain randomizer feedback values that are unusable. The system has been designed to make sure that enough randomizer feedback are simultaneously evaluated so that a usable feedback is always available. For purposes of evaluation, the system evaluates the number of paired, tripled, and quadrupled output vectors in determining which randomizer feedback to use, as well as to determine when a randomizer feedback should be discarded.
For a given number of output states, a given number of input data patterns, and a given number of multiple outputs, it is possible to determine the probability that any specific randomizer feedback maps the input data patterns into a usable set of output states. The analysis below uses a total of 10000 input data patterns, 65536 (2ˆ16) possible output states, and 1024 (2ˆ10) possible multiple outputs. For this scenario, it can be calculated that any possible randomizer feedback has an 95% chance of producing a usable mapping of the input data patterns. By using a set of eight possible randomizer feedback, the odds of having a usuable mapping are 99.9999999961%.
The presently preferred embodiment of the system can use a total of 128 (2ˆ7) possible randomizer feedback. When one or more of the eight randomizer feedback whose mapping has been evaluated becomes unusable, the system can use one of the remaining (128−8)=120 randomizer feedback. It should be noted, that from a practical standpoint, new feedback paths can be swapped in while the input data patterns are loaded into the system.
Secondary Randomizer Feedback Selection Probabilities
The secondary randomizer differentiates between input data patterns that have been mapped to the same output value. The calculations for the odds of having a usable secondary randomizer feedback value are shown below. The probability analysis for this operation is much different than for the primary randomizer because, in this case, it is only necessary to be sure that the entries in each multiple are different from each other.
Randomizer Scaling for Additional Inputs
If the user wants to support additional inputs using the system, it is possible to scale up the size of the randomizers to achieve functionality. If the number of inputs were scaled to 40,000 input data patterns, the number of outputs could be increased to 262,144 outputs, and the number of multiple outputs could be increased to 4096. This would require an increase in the length of the primary and secondary randomizers to 18 bits in length. In addition, the memory requirements for storing the randomizer feedback mappings would increase by a factor of four.
System Overview
The primary method of operation for the system is used for ultra-high speed packet classification of optical data that has been framed into a serial data stream. This method of operation is referred to herein as the serial mode of classification. The invention provides packet classification at the serial data stream level, as opposed to doing this after data has been put into a parallel format and transferred into a network processor system. This feature of the invention allows the system to produce extremely fast characterization in a predictable timeframe that exceed anything done in a traditional parallel form such as the example previously shown of a CAM based system. The serial mode of classification requires modifications to a standard data framer part in addition to the other components that make up the system.
A secondary method of operation for the system provides fast packet classification of data packets that have already been stored in memory as successive parallel words of data. This method of operation is referred to herein as the parallel mode of classification. The purpose of this method of operation is to provide a legacy mode of operation that can support on or more ports that may or may not have modified data framers supporting the system. A design in which all ports have modified data framers supporting the system does not require use of the parallel mode of classification.
For the primary mode of operation, the system incorporates unique features into a traditional optical data framer chip, and relies on a complex ASIC to permit the user to differentiate between thousands of different patterns at light speed. One reason for the general purpose system is to eliminate the need for costly and power consumptive content addressable memory systems, or customer pattern specific ASICs, to perform network packet classification. The system operates on a principle of adaptive programmable randomization to permit the differentiation between the input vectors to be made. The system can dramatically reduce the processing burden required by high-speed optical routers or switches.
The presently preferred implementation of the herein disclosed architecture supports identification of up to 10,000 distinct input vectors that can be up to 1024 bits in length. Both the number of vectors and length of each vector can be modified should the situation require. Increasing the number of vectors beyond 10,000 would require increasing the length of the programmable randomizers and the width of the SRAM memory used by the custom ASIC. Increasing the length of the inputs results in a fairly linear increase in the overall size of the custom ASIC and would require slight modifications to the data framer ASIC.
Novel Data Framer
The novel data framer modifications are made to the serial data stream of a standard data framer chip. If a descrambling function is done, it is important that the data framer modifications be done after the descrambling function. The serial data stream must be the same as the parallel data that is going to be passed to the network processor. It is also important that the data framer be able to access the data framer outputs that byte align the start of the packet, as well as count the number of bytes received.
The next change to a standard data framer part involves the addition of a randomizer enable control block 33. The clock to the primary and secondary randomizers are gated ON and OFF by the enable randomizer signal that is generated by this block. The enable randomizer signal is turned on at the start of a packet.
An optional implementation allows the data to the randomizers to be programmably turned OFF (Set to a 0) and then turned ON again to allow blanking out a portion of every packet received. This is done with the gate randomizer signal. The implementation shown in
The optional masking control block 34 in the data framer allows programmable, sequential, user controlled masking of groups of user defined bits. The implementation that is shown allows four different 32-bit wide blocks of data to be captured for masking purposes. The user is able to configure with MASK0_START[4:0] to MASK3_START[4:0] four starting locations for 32-bit words to be sampled. The masking control block takes these four START values, in addition to the BIT_COUNT[9:0] register which is a count of the number of bits from the start of the packet, to determine when to sample the MASKING data.
The PAR_DATA[31:0] is parallel data from the serial to parallel converter 35 which is implemented within any data framer. This PAR_DATA[31:0] is sampled at the appropriate times to generate MASK0DATA[31:0] to MASK3DATA[31:0] which are four 32-bit wide masking registers that are associated with the packet, and can be read by the custom ASIC when they become available. The number of MASK registers can be modified with changes to the data framer as well as the custom ASIC should an application warrant this being done.
The output register synchronization and queue 36 assures that the primary randomizer, secondary randomizer, feedback registers, and MASKING registers are stored for each packet. If a queue of packets is implemented in the data framer, then these registers must be similarly queued so that they are associated with the appropriate data packets. Ideally, these registers should be available to the custom ASIC as soon as the number of bits identified by the STOP register are received, so that they can be used to determine the appropriate input pattern match.
Primary and Secondary Randomizers
The primary and secondary randomizers are two equivalent circuits, and are designed so that their physical layout can be done so as to minimize gate propagation delays and hence to permit extremely fast operating speeds. These randomizers are expected to easily operate at 10 Gbits/sec, and as technology improves in the coming years, the step up to 40 Gbit/sec should be possible.
The randomizers for the system are sixteen bits in length. They are constructed using sixteen stages of D Flip/Flops in a serial shift register as shown in
A global Clear signal is also applied to the randomizers. This signal is used to initialize the randomizers to a known all 0's state between packet receptions. The Clear signal is not time critical so long as it is applied after the randomizer values have been latched, and prior to the start of reception of the next data packet.
The feedback for the randomizers is structured as an XOR tree that potentially sums up all sixteen of the serial shift register bits and adds them to the Input data bit. For the case of a 16-bit randomizer, this translates to five levels of exclusive- or logic. The randomizer feedback are turned on and off with the PRIMFB[14:0] and SECFB[14:0] registers. The last shift register bit is always applied to the feedback network, and hence the reason that there are only fifteen feedback selection bits. Each feedback bit is used to gate the output of its corresponding serial shift register stage using a two input AND gate. The PRIMFB[14:0] and SECFB[14:0] registers should only be written to between receptions of data packets. Changing these values does not affect the actual primary randomizer or secondary randomizer outputs if it is done while the randomizer clocks are gated off.
Enable and ON/OFF Circuitry
A key addition to a standard data framer is the generation of an Enable signal for the primary and secondary randomizers. At a minimum, this signal should enable the randomizers when a packet reception begins, and should disable the randomizers when a specified number of bits have been received. The optional ON/OFF circuitry in the custom ASIC handles gating off large blocks of data that are not desired to be analyzed, occurring in the middle of data fields the user wants to analyze with the system. This could consist of cyclic redundancy checks or other fixed blocks of data in the received packet that are not pertinent to the packet classification process. The purpose of the ON/OFF circuitry is to provide greater flexibility for use of the variable MASKING bits.
In the optional case where ON/OFF circuitry is implemented, additional registers and comparators are required. For each long blanking period that is desired, an additional two 10-bit registers 93, 94 are needed to store the ON and OFF bit periods. In addition, two 10-bit comparators 95, 96 are required to compare these ON and OFF values to the actual bit count. The outputs of all of the comparators drive the enable state machine. To increase the number of ON/OFF periods that are blanked requires additional logic and power consumption. The enable circuitry comparators are operating at the full clock rate of the system because they are driven by the bit count value.
Programmable Masking Circuitry
The ability to execute programmable masking in the data framer is entirely optional. In the world of content addressable memories, this feature is akin to a ternary CAM having three levels, i.e. one, zero, and don't care. The programmable masking circuitry permits blocks of data to be captured so that they can be programmably masked in the final output result. To reduce power consumption and chip area, the programmable mask registers have been defined to be on even 32-bit boundaries, and to be 32 bits in length. The size, boundary, and length of these registers could all be adjusted based on user requirements. The programmable masking circuitry performs a function similar to a logic analyzer, by capturing blocks of data at the appropriate time, and then permitting them to be analyzed in the future in the custom ASIC.
In the implementation shown in
The masking data are captured in a set of 32-bit registers labeled MASKDATA0 103, MASKDATA1 104, MASKDATA2 105, and MASKDATA3 (not shown). These registers are clocked when the Bit Count equals the value stored in the respective Mask_Start register. The data input to the MASKDATA registers is from the serial to parallel data converter that exists in a standard data framer part. Once masking data are captured, they must be read from the output registers prior to the next packet being received.
Output Register Synchronization and Queue
It is critical that the primary and secondary randomizer values be synchronized to the feedback values that were used in determining them (see
Custom ASIC Implementation
The preferred embodiment of the custom ASIC has four primary interfaces to the outside world, i.e. the microprocessor interface 41 is used to communicate to a host processor system, the DRAM interface 42 is used to communicate with either a standalone DRAM or shared dual port DRAM for storing data patterns to be matched, the SRAM interface 43 is used to communicate with a dedicated SRAM that contains mappings for various primary and secondary randomizer settings, and the interface 44 is used to communicate with the modified data framer part. Internally, the custom ASIC handles mapping user inputs into values that are stored in SRAM. Individual masking and mapping optimization functions are performed internal to the ASIC.
The following discussion views the custom ASIC from a data flow perspective:
For the custom ASIC to start it's process, the user is required to make various pointer and register configurations (see below). Next, the user must load input patterns into the custom ASIC. These patterns are made up of a combination of data and, in some cases, masking steps that should be done on the data in a potentially sequential fashion. For instance, the custom ASIC permits mapping a range of data values to a single mask step output, and when that mask step output is reached, additional programmable masking or verification can be executed. Input patterns are handled by the input manager control and state machines function 45 where they are directed into the input register 46. They are also loaded into external DRAM for use in cases where an equation mapping is discarded, and a new mapping must be generated from scratch. The input register is 1024 bits long to permit patterns up to 1024 bits to be analyzed. This length could be adjusted up or down in embodiments, and would not affect the data framer ASIC. Input data are manipulated in 32-bit words, and masking is permitted for 32-bit lengths on 32-bit word boundaries. Masking information is stored in conjunction with the input data so that an entire input pattern can be regenerated should new equation mappings need to be generated.
When the 1024 bit input register contains a complete data pattern, and when all masking information is loaded in the masking and enabling logic section 47, it is possible to use the 1024 equation mapper 48 to generate a related randomizer value. The masking and enabling logic 47 is responsible for enabling only those input register bits that the user is interested in examining. For instance, if the user is interested in reviewing only the first 200 bits of an incoming data packet, the last 824 bits are blanked out by the masking and enabling logic, and are not sent into the primary and secondary randomizers in the data framer. The masking function of the masking and enabling logic allows data bits within the input pattern to always be masked out and not analyzed, as well as permits bits to be masked selectively in a user programmable sequence.
The equation mapper 48 permits a randomizer value to be calculated in a single cycle. This is done by implementing the mapper in pure logic gates in hardware. It calculates randomizer mappings for 128 different equations simultaneously to permit extremely fast calculations and data storage of randomizer values generated by different randomizer equations. This feature permits the custom ASIC to adjust adaptively to select optimal randomizer settings based on the input and masking patterns that have been applied to the part. It allows many randomizer mappings to be stored at one time without adversely affecting the setup time for the part, and makes the action of generating entirely new mappings possible in the case where some of the mappings stored on the custom ASIC are no longer usable due to excessive randomizer output duplication.
The mapper multiplexer 49 is implemented in hardware to allow immediate selection between each of the possible 128 randomizer outputs associated with each of the 128 possible randomizer equations. This function preferably consists of sixteen individual 128:1 multiplexers, where there is a multiplexer associated with each of the sixteen output bits of the randomizer. If the length of the randomizer is increased, the size of this multiplexer increases in a linear fashion with the number of bits in the randomizer output. The number of equations used in the equation mapper and the mapper multiplexer can be adjusted up or down with a direct impact on the number of gates used in their implementation.
The mapper storage control and storage state machine 54 is responsible for saving and retrieving values from the various equation mapping tables. The state machines in this block handle determining whether the present location pointed to by a primary randomizer value contains 0, 1, 2, 3, or 4 entries. This state machine is responsible for handling the creation and destruction of multiple entries, and the adjustment of the multiple entry tables that dictate those entries that are used. This function works directly with the external SRAM in storing and retrieving these values. This function works closely with the masking engine, especially in those cases where mask steps are involved.
The masking engine 51 handles all aspects of permitting the user to setup sequential masking operations. This function takes user inputs that setup which, if any, 32-bit blocks of data are operated on in a masking fashion. The user is permitted to signify data bits as always being masked, or as selectively being masked when certain circumstances arise. Once the masking engine knows what bits may be masked, it is responsible for calculating the effect of each bit on the output vector. These effects are referred to as masking impact bits. To execute this operation, the masking engine injects a single “1” into the input vector stream at each bits location. The output primary randomizer value shows the impact of this input on the output vector. This value is then used by the masking engine, along with the masking data that is received from the data framer ASIC, to remove the effects of the masked data when a packet is received. The masking functions used in the system are described in greater detail below.
The time accelerator 52 is responsible for re-mapping a received randomizer value to generate the randomizer value that would have been received if zero values had been clocked into the randomizer for a fixed number of cycles after the received randomizer value was captured. This function is described below, and permits a flexible and fast way to skip forward to the end of the 1024 input bits that are always used to calculate randomizer values. As an example, if the user wishes to analyze only 200 bits of data, while the novel custom ASIC always calculates using an input vector length of 1024 bits, the time accelerator block generates the effect of having 824 trailing 0's shifted into the randomizer after the data of interest. This method reduces time latency, and permits the randomizers to be turned off much sooner to reduce power consumption.
The mapper engine, statistics and state machine 50 is responsible for determining the equations to be used by the system, and to determine when equations are no longer usable and need to be replaced. This function maintains statistics on all equation mappings that are maintained in memory. Using these statistics, it selects the best mapping and sends it to the data framer ASIC.
The master control block 53 is responsible for initializing the entire system at startup. It also handles communicating with other control blocks to execute system wide functions such as a hard reset.
SRAM Memory
Dedicated SRAM is required for the custom ASIC. The size of the SRAM memory is dictated by the size of the primary randomizer/secondary randomizer words, and the number of equations that are permitted. The speed of the SRAM is critical because it sets the performance of the system. As an example, a 16-bit primary randomizer value, in conjunction with a 16 bit secondary randomizer value, dictates that the core of the primary randomizer lookup table be 2*2ˆ16=131,072 entries long and sixteen bits wide. In addition, a multiple entry table of 1024 entries contains slightly more than 8,096 locations that are sixteen bits wide. In this example, it takes a total of approximately 140,000×16 bits of SRAM storage per equation. A 512K×6 SRAM handles approximately three and one-half equations worth of data.
DRAM Memory
A dual port DRAM could be used for the DRAM required in the system. The lowest cost configuration is a synchronous DRAM. The fact that the custom ASIC is the only device talking to the DRAM, unless a dual port approach is used, permits a purely synchronous approach to be simply implemented. The speed of the DRAM memory directly impacts how fast new input vectors can be stored in the system. The microprocessor interface to the ASIC is a 32-bit interface. If the DRAM is sixteen bits wide, then it's interface must be twice as fast. Burst operation is very critical because an entire input could be stored away at one time.
System Benefits
Fastest Packet Processing Possible.
As soon as the last serial data bit that is under analysis has been received, the system's primary randomizer, secondary randomizer and masking bit values are available from the data framer ASIC. These words can be read into the custom
ASIC using high speed parallel 32 bit transfers (taking <5 nsec). Once these data are transferred to the novel custom ASIC, the time to match a pattern is a function of three variables, i.e. the number of masking operations that the user wants to perform, the speed of the SRAM used by the custom ASIC, and the distribution of multiple outputs within the search that is performed.
Multiple outputs are distributed in a probabilistic fashion throughout the search process. The odds described below assume that there are up to 968 paired outputs, up to 50 tripled outputs, and up to six quadrupled outputs within a 10000 input data pattern space. In this scenario, the odds of a pair are less than 2*968/10000=19.4%, the odds of a triple are less than 3*50/10000=1.5%, and the odds of a quadruple are less than 4*6/10000=0.24%. These are extremely conservative bounds because, in the example, all of the probability distribution is for cases that are less than or equal to the 968 pair, 50 triple, and six quadruple scenario. When a pair, triple, or quadruple output is hit, the system uses the secondary randomizer value to differentiate between the different input data patterns.
The speed of the SRAM used by the custom ASIC dictates the time associated with sequential SRAM access to determine the appropriate input data pattern. In addition, each sequential masking step, excluding a fixed masking of bits which does not entail SRAM cycle overhead, results in an additional search through SRAM memory.
The search time for a pairs, triples, and quadruples is shown in Tables I-L below. Each of these tables lists the average and maximum time to complete a search.
In calculating the average search steps per mask step, the previous probabilities for multiple outputs can be used.
The SRAM speed, could be increased to improve the overall search time. The flexible masking steps are an additional feature that require additional searches in CAM solutions. It should be remembered that there is no additional overhead for a fixed masking of data bits. The search time provided by this embodiment of the invention is approximately five times faster than comparable CAM implementations, and is on par with the search time of a hardwired ASIC2 approach by Juniper Networks. The highlights with this approach are that it is extremely deep in it's search, it is flexible in terms of allowing multiple masking operations, and it has the lowest latency of any solution.
Lowest Possible Search Latency
The system has the lowest possible packet search latency that is presently known. This occurs due to the fact that the system operates on a serial data stream prior to any parallel operations of transferring the data into memory. Alternative CAM or ASIC based systems must transfer all appropriate data bytes from a data framer into either a processor memory or directly into a CAM or ASIC solution for evaluation. The time required to perform this function adds latency to the packet through either a router or switch product. Reducing this latency helps guarantee that higher level protocols do not time out as packets are routed over an entire link.
The system can produce a guaranteed upper bound on the time required to evaluate a received packet after the packet is received. This deterministic delay could be useful in an alternative system where optical routing is performed using a detection system and an optical delay line.
The optical router architecture shown in
In addition to the physical challenges of the delay line, the signal protocol issues would have to be developed. All optical switches being developed today setup a connection and leave it there for a long period of time. This means that their switching time is not critical, and the connection is bi-directional.
Lower System Power Consumption
Power consumption is a critical parameter on line cards and in other networking gear. Reducing power consumption can have a ripple effect on overall system costs, and can also provide more degrees of freedom for the board and system designer. The system has been designed to take the most processing intensive burden away from the network processor. This permits designers to react by reducing clock speeds or decreasing memory bandwidths and achieve equivalent performance. Either of these choices reduces power consumption and cost.
The system has been designed for minimal power consumption. The data framer primary and secondary randomizers operate at full clock speeds, but are only driven during the portion of a receive data packet that is being evaluated. The design of these randomizers minimizes their power consumption. The remainder of the data framer consists of configuration or sampling registers that are read or written only once per packet, and hence draw very little power.
As with the data framer, the custom ASIC has also been designed for minimum power consumption. The custom ASIC and associated memory system are very active during configuration and draw higher power levels at that time. Once configuration is complete, the custom ASIC is fairly static. It executes SRAM cycles when primary and secondary randomizer words are read from the data framer, but these are limited in duration. Otherwise, the custom ASIC operates with little dynamic power consumption. In contrast to the system, both a prior art custom ASIC and a CAM system require extensive power consumption while they are active.
Flexible and Programmable Masking on the Fly.
Typical CAM based search engines permit global masking and bit specific masking of any bit in the data pattern. Generally, there are a limited number of bit patterns where this capability is required. The flexible and pre-programmable masking sequences that the system permits, offload additional processing work from the network processor. Based on the type of packet being received, the system can process the packet with different sequential masking operations.
Using the System on a Router Backbone
In the interest of cost savings, the system permits multiple data framers to be operated from a single custom ASIC. Many systems, such as the Juniper ASIC2 and prior art CAM based systems, can operate with the packet classification processing done on a network router backbone as opposed to on each channel. In the system, each input channel requires a data framer ASIC, but a performance tradeoff allows a single custom ASIC to be used for N data framers.
In terms of implementation, the data framer interfaces are preferably all in parallel so that a single custom ASIC 25a could write to each data framer 22 at once. This entails setting up the randomizer feedback values. The randomizer outputs, masking registers, and feedback values used to analyze each packet are added on to each received packet. These packets are sent through each network processor 61 onto the router backbone. On the router backbone, a main router processor 62 is responsible for queuing the packets, stripping the data off and routing it into the custom ASIC, and using the custom ASIC outputs to determine how to route the packet.
Unsynchronized Data Pattern Extraction Capability
The architecture of the system provides a unique opportunity to perform ultra-high speed searching for many unsynchronized data patterns in a stream of high speed serial data. By the term unsynchronized data pattern, what is meant is a data pattern where the start of the pattern is not known, and the pattern is embedded in a stream of data. In the case of searching a data stream for specific text or other patterns, the system allows up to 10000 patterns of up to 1024 bits to be searched simultaneously in a rapid fashion. In the future of genetic research, if it was possible to take a sample of genetic information, and break it down into it's sequences while doing this in a brute force fashion on all chromosomes simultaneously, the following based approach would permit all of the genetic data to be scanned rapidly for matches with up to 10000 genetic sequences. In this way, an extremely rapid analysis of a genetic fingerprint for an individual could be made. Each base in a DNA strand can be represented with two bits of data, so the ASIC could find strands that are up to 512 bases in length. The length could be increased beyond 512 bases with a simple redesign to the novel custom ASIC.
The system for scanning random data for a match can be implemented using multiple novel style circuits as shown in
The single bit offset between novel data framers also means that one of the data framers produces an output each bit period. The n:1 multiplexer routes the correct data framer output to the novel custom ASIC 25 during each bit period. During this time, the primary and secondary randomizer outputs and any masking outputs must be transferred into the novel custom ASIC and used there. The serial speed of this system is limited by the search time of the system. The current worst case pattern match takes six SRAM cycles for the case of a quadruple output match. Using a worst case of 50 nsec for a match, the serial data rate can be up to 20 Mbits/sec if only fixed masking operations are used. A serial data rate of 20 Mbits/sec allows the data framer ASIC circuitry to be implemented easily in silicon.
One issue that affects the asynchronous data stream much more than the optical networking situation is the case of false detection. If sixteen bit primary and secondary randomizers are used, there is a one in 65,536 chance that a valid primary randomizer number has random data generate a matching secondary randomizer value. There are a variety of ways in which this can be handled. A first method is to check the data always to make sure that it was a match; a second method is to increase the size of the memory and hence reduce the odds of both valid primary and secondary values; and a third method is to use an offset pattern check to make sure that two successive readings point to the same value. The offset pattern check can be used if less than ¼ of the possible data patterns are being searched. In this case, the length of the search pattern can be increased by one bit. Every desired pattern is prepended by 0, prepended by 1, appended by 0, and appended by 1 to generate four new patterns that are each one bit longer than the desired pattern. When the system receives a match for a specific data pattern, there must always be a match for the very same specific data pattern on the very next bit period. If this does not occur, there was a false detection. This approach decreases the odds of a false detection by a factor of four billion (2ˆ16)ˆ2.
Another issue that could arise in searching the genome is the fact that input data patterns are of various lengths based on the gene. This requires there to be additional masking registers to permit a wide range in data lengths. As an example, if strands of 26, 52, 64, 78, 84, 100, 160, 220, and 330 bits in length were to be checked, the masking would have to be done sequentially, and would reduce the overall throughput. All strands would be verified for the first 26 bits, and the result would direct towards measuring the next (52−26)=26 bits for all cases that were not the 26 bit strand. At that point, the (52−26)=26 bits would be verified, and the result would direct towards measuring the next (64−52)=12 bits for those cases that were not the 52 bit strand. This would extend on until all the steps were taken.
Normally, if ten masking steps were used, the throughput of the system would be reduced by a factor of ten, but the throughput reduction could be greatly minimized through a variety of methods. In the case of genetic pattern matching, the frequency of cases where more than one mask step would be taken is extremely small, and usually would occur only if a valid pattern was correctly bit aligned. As a result, the system must handle extremely rare cases where multiple masking steps may be taken. This could be done efficiently by using a FIFO (first in first out) buffer in front of the novel custom ASIC to buffer the incoming randomizer values to handle the multiple masking cases, or by shutting off the serial data clock until a check has been completed. The important criteria here is that the most efficient searches occur when all of the data patterns to be checked are of the same length.
The novel data system opens up exciting opportunities in the arena of data mining and pattern matching. Table M identifies the time that this sort of an asynchronous data pattern detection system would take to search through the entire-Human Genome. Printed text is not a good example of an asynchronous data stream because it is already framed at the character level, and can be easily translated to be framed at the word and/or sentence level.
Features of the System
Highly Specialized Programmable Randomizer in a Serial Data Path
The programmable randomizer in the system is unique in a number of ways ranging from the method of programmability, the random nature of it's outputs, the simplicity of making predictive calculations as to it's value over a long stream of inputs, and the ease of predicting the affects of any specific input on the output state to effect rapid masking calculations.
Linear feedback shift registers have been used for decades in a variety of applications including cyclical redundancy check generators (CRCs), pseudo random bit sequence generators (PRBSs), and data scramblers. In most of the known applications, the feedback taps on these shift registers are selected at the point of design to maximize their randomization features. The pseudo random bit sequence generator and data scrambler applications use a linear feedback shift register to produce a serial data pattern that is as random as possible. The cyclical redundancy check generators operate on a serial data stream, and produce a parallel word at the completion of reception that is used to verify that the correct data stream has been received. In this respect, the system's use of a linear feedback shift register is similar to the application of a cyclical redundancy check generator.
The system differs dramatically from a CRC generator in the respect that it has programmable taps that allow 2ˆ(n-1) different programmable feedback to be selected instead of a single fixed feedback. The “n-1” term arises because the final “nth” stage of the shift register must always be fedback in the preferred implementation. Further, the feedback are setup to exclusively-or the serial data with any combination of (n-1) shift register stages. The exclusive-or feedback allows a fixed tree of exclusive or gates to be used to exclusively-or all of the selected feedback outputs. Each output is gated with an “and” gate to control programmably whether or not it is used by the exclusive-or tree. A key feature of this approach is that it is extremely compact and easy to program. This dramatically reduces die size and power consumption because the circuitry is meant to operate at extremely high serial data rates, and the more logic that is toggling at these data rates, the more power that is consumed.
A purpose of the randomizers is to differentiate between input data patterns, and not to provide the most random possible pattern. What is more critical, is that each input affects outputs differently when switching between various feedback mechanisms. This orthogonality between how different outputs are impacted by each input, through a range of equation feedback, permits general random probability theory to be used in calculating output distributions.
The design and choice of the randomizer is critical to the ability to predict quickly in hardware the mapping from an input data stream to an output randomizer value. The sole use of exclusive-or gates in the feedback mechanism, as opposed to other forms of logic such as “AND,” “NAND,” “OR,” or “NOR” gates greatly simplifies the predictive ability in hardware. In addition, the use of exclusive-or gates makes it possible to single out each individual input, and identify it's effect on each bit of the output pattern. This feature is critical for permitting simple masking to be performed. Finally, this approach is critical to permitting variable length packets to be easily analyzed. This is possible by the fact that any bit that is beyond the length of data to be evaluated, can be forced to a 0 in the analysis hardware tree, and the bit is effectively eliminated.
Single Cycle Hardware Calculation of Randomizer Values
The system operates on the ability to pick the best feedback mechanism for mapping input data patterns into output vectors in a rapid fashion. This requires that the system must be able to calculate output vectors in an extremely rapid fashion for a number of feedback selections. The system uses a hardware tree of exclusive-or gates to calculate output vectors for a large number of feedback selections (nominally 1000 feedback selections) in a single clock cycle. This is possible to implement by analyzing the effects of each and every input bit on the output pattern using a computer program in advance. Once these computer mappings are performed, they can be merged together for all of the possible feedback so that as many logic gates as possible are re-used. Finally, an output multiplexer allows the appropriate output vector to be selected for the feedback mechanism that is being evaluated.
Randomizer Gating Implementation
The primary and secondary randomizers in the system are gated on and off for three specific reasons, i.e. at the initiation of reception of a data packet they are turned on, at the completion of the reception of the data bits of interest they are turned off, and during any period in which the user always desires to blank out analysis of a sequence of data bits they are turned off and then back on at the end of the sequence. This ON/OFF gating of the randomizers, through user programmable control, is a unique feature to the system. This ON/OFF gating capability works in tandem with the single cycle hardware calculation of randomizers circuitry, where data bits during any “OFF” period are forced to a “0” from the perspective of randomizer calculations.
Programmable Masking Architecture
In CAM or ASIC implementations of packet classification systems, data bits can be masked off by essentially not considering them in a bit by bit comparison. The masking capability in the system is very unique. Instead of the traditional approach of blanking out or ignoring a data bit to mask it, the system calculates the effect of each data bit that the user desires to mask on the randomizer outputs, captures the actual masked data bits, and then effectively subtracts the masked data bits out of the randomizer output result using specialized hardware. The data framer incorporates the circuitry to capture the masked data bits, and the custom ASIC incorporates extensive circuitry to calculate the effects of each masked data bit on the output predictively, as well as the circuitry to subtract out the effects of the masked data bits.
Sequential Masking Capability
In addition to the ability to mask out data bits in a received packet stream, the system allows the user to setup programmed sequential masking of data bits where initial pattern matches can automatically drive subsequent masking operations without user intervention during the processing of the randomizer outputs. This sequential masking capability is normally done through processor intervention based on the outputs of various decisions. The system permits this decision driven masking to be setup proactively in advance of a packets reception. This reduces the overhead required of a network processor, and reduces the latency in making a routing decision.
Adaptive Randomizer Feedback Analysis and Selection
The architecture revolves around the ability to develop and maintain information regarding a number of randomizer feedback at any given instant in time, and to pick the best randomizer feedback adaptively using established criteria. The custom ASIC manages the memory (SRAM) tables associated with each feedback selections mappings. In addition, counters for the number of pairs, triples, quadruples, and overflow output vectors are maintained in hardware. These counters are used to identify the randomizer feedback that is least likely to need to be changed, as well as any randomizer feedback that are no longer viable. The evaluation and swapping of randomizer feedback is a real time and adaptive process based on the latest information as to the input data patterns being evaluated.
Input Manager Control and State Machines
In the preferred system, the user is allowed to provide up to 10000 different inputs that are up to 1024 bits long. This memory is a maximum of 10,240,000 bits of storage. The system returns a pointer to the data input that matches the incoming data stream. Therefore, the user should store the data in sequential locations. A single valid bit for each input is used to signify that the input data pattern is valid and should be used by the system. The external memory system is specified as using 32-bit wide memory. Therefore, it requires 32 memory accesses to read a maximum length data input. The user can write an input, and then activate that input by writing to it's associated valid bit.
External Memory Input data and Input Valid Structures
The purpose of the external DRAM memory storage is to store input data patterns and masking information for potential future needs. If certain randomizer feedback must be eliminated due to excessive multiple structures or multiple structure overflows, it is necessary for the system to read all of the input data patterns back into the system from memory so that new patterns can be evaluated. Another reason for storing the input data is to permit users to read back any input data pattern that is being used in the custom ASIC system to evaluate the pattern, or to modify it to create another input data pattern.
The user can define the location in the external memory where the custom ASIC stores the input table. This table is called the INPUT_DATA array, and it's starting location is pointed to by the 32-bit pointer INPUT_DATA BASE that is located on the custom ASIC. Thirty-two bits of width allows for an addressable memory size of 4.29 G.
The user must define the number of bits of data that each input data pattern contain, and this value is stored in INPUT_DATA_LENGTH that is located on the custom ASIC. This value can be in one bit increments, but from a storage perspective, input data patterns are stored in 32-bit words. Any data bits beyond this length value are masked out. In the case of an exclusive-or function, this is done by masking the input to a 0. For the custom ASIC, the maximum valid value is 1024 bits. In embodiments this value could be extended in length.
In addition to the input data pattern, the system must know what masking is required for the specified input. This information is referred to as the present masking step. Any bits in the pattern that are masked are ignored from the perspective of the stored input data pattern. If a masking step is performed after the present input data pattern with optional masking is evaluated, the next masking step must be specified. These two masking steps are stored after all of the input data pattern data words and are associated with the specific input data pattern (see Table N below).
The user must also signify which inputs are valid. This is done with the INPUT_VALID array, where a single bit is used to signify the validity of each user input. To conserve on data, this array is a 10000x1 bit array that can be accessed in 32-bit reads/writes. The first 32-bit location in the array signifies the status of inputs 0-31. The second 32 bit location in the array signifies the status of inputs 32-63, etc. Coverage of 10,000 Input data patterns requires a total of 313 32-bit words. The INPUT_VALID BASE is used to indicate the location of the array that contains the valid status for each input location (see Table O below).
The shaded area includes valid entries for Inputs 10,000 to 10,015 which are unspecified.
The INPUT_DATA and INPUT_VALID information is all stored in external DRAM memory. Table P below describes the location of these arrays.
Registers Associated with Input Data Structures
The following registers are located on the custom ASIC, and are used to write and clear user input data patterns located in the external INPUT_DATA array. In addition, they are used to modify the INPUT_VALID array appropriately. The INPUT_DATA_BASE and INPUT_VALID_BASE registers store the base addresses of the external structures as described herein, and the INPUT_DATA_LENGTH register stores the length of each input as previously described. The INPUT_DATA_WORD_COUNT is used as a counter to point to the individual words in the INPUT_DATA array. This value is automatically incremented after each word is written to, or read from, the INPUT_DATA array. The INPUT_DATA_NUMBER refers to the input data location that is being operated on, and falls in the range of 0 to 9,999. This location must be written to prior to operating on the INPUT_DATA array, and is stored in the USER_INPUT_DATA_NUMBER location. The INPUT_AUTO_LOCATION register is used by the system to identify the next available user input that has not been setup as valid. If the user wishes to use this input for the next Input data Pattern, the value should be read, and then written into the USER_INPUT_DATA_NUMBER register.
Within the input manager control and state machine, the necessary information regarding masking regards the masking required for the described input, and whether the randomizer results drive an additional mask step, or whether they are a final result. A later section on masking describes in detail how the masking is handled and calculated. The PRESENT_MASK_STEP value stores the masking step to be used on the present data inputs. Any bits that are covered by this masking operation are disregarded with respect to their storage in the INPUT_DATA array. The NEXT_MASK_STEP value dictates whether there is an additional masking step. A NEXT_MASK_STEP value of 0 indicates that this is a final value, and when reached, the user receives the INPUT_NUMBER for the corresponding INPUT_DATA pattern that has been matched. The two MASK registers are stored automatically by the system after all of the input data words have been loaded into DRAM.
In addition to the aforementioned registers, there are a group of registers within the input manager control and state machines (see Table Q below) that are used for internal operations. INPUT_STRUCT_PTR is a general purpose pointer used to manipulate data. INPUT_STRUCT_VALUE is a general purpose register used to store either INPUT_DATA or INPUT_VALID information. INPUT_VALID_ENCODE is used to store the result of a 32 to 1 priority encoder, and the INPUT_CONTROL_REG is described below.
INPUT_CONTROL_REG Details
The INPUT_CONTROL_REG (see Table R below) contains all of the control bits associated with modifications to the INPUT_DATA array and the INPUT_VALID array. These bits include those necessary to Initialize the INPUT_VALID array, to write new INPUT_DATA, and to read INPUT_DATA. The special AUTO_LOCATION feature is used to permit the system to determine the next available INPUT_DATA location. This is an option that is permitted in cases where the user does not want to keep track of this process in a tight fashion.
Inputs and Equation Correlation
It is critical that there be correlation between the inputs that are stored and the ones that have been mapped by various equations. As inputs are written or cleared, they are always calculated for all active equations at that time. The equation update generator uses an EQUATION_INPUT_UPDATE_PTR to cycle through all of the valid inputs. If a new input is stored in the middle of this process, or an input is cleared, the EQUATION_INPUT_UPDATE can be temporarily discontinued to allow that to occur. When equation updates are being made, it is necessary to initialize (clear) out the entire primary and secondary randomizer tables.
Processes Associated with Input Control
The processes shown in Table S below are used to manipulate the input array and the input valid databases.
Input Register Arbitration and DRAM Access Arbitration
There are five processes that require access to the DRAM. Two of these processes, USER_INPUT_WR_LOAD, and USER_INPUT_CLEAR, also load the input register The user should be accessing only one of the three user/external processes at a time because it is the user's responsibility to finish one action prior to starting another. From the perspective of the 1024-bit input register, there are three processes that result in it being loaded: USER_INPUT_WR_LOAD, USER_INPUT_CLEAR and SYS_INPUT_LOAD. The SYS_INPUT_LOAD and USER_SYS_INPUT_VALID procedures can be initiated by the system, as part of a full equation swap, and are less critical than an immediate change by the user with either the USER_INPUT_WR_LOAD or the USER_INPUT_CLEAR processes (see Table T below).
The DRAM memory can be accessed by a number of processes described above. It is necessary to have arbitration between these processes, so that any two do not conflict over the use of internal registers. The weights in the Table U below are meant to show how often one of the following actions should be taken when two actions are being requested. The real issue here is the relative weights. For example, a USER_INPUT_WR_LOAD occurs eight times for every USER_INPUT_LOAD if both are continually being requested.
Inter-Block Communication
It is necessary for the mapper engine, statistics, and equation state machine to be able to read input data values from DRAM into the input registers so that they can be mapped and stored. This is important for the equation update process.
The I/O ready bit in the input control register is used to hold-off read or writes to the custom ASIC until the proper internal function has been completed. In addition, the I/O ready bit is used to handle the larger issue of holding off the user while the system uses the input registers to recalculate an equation. A round robin arbiter is used to make sure that the system and the user get alternating preference on occasions where a conflict exists. As the user issues a command, the system determines whether the mapper engine, statistics, and equation state machine is also requesting access. If the internal mapper . . . ″ block has priority, the I/O ready bit is kept at a 0 level until the internal operation completes (see Table V below).
1024-Bit Input Registers
There are four sources of data that need to be able to drive the equation mapper, i.e. a new user input that should be mapped and stored for future use, a previously stored user input that should be brought back for analysis with a new equation, special bit patterns required for input data masking analysis, and parallel data that is received and requires packet classification. The discussion above described how arbitration between new user inputs and previously stored user inputs is performed. Neither of these functions is time critical, but either requires significant overhead in setup and should not be interrupted prior to completion. The special bit patterns required for input data masking analysis are described herein, and they are independently generated and multiplexed into the source of the equation mapper. Finally, parallel data that are received and require packet classification have a time critical component. Due to the time criticality for parallel classification, an independent set of registers is used to store this data. This permits decisions to be made rapidly, and results in minimal disruption to the new user input and previous user input analysis that may be occurring concurrently.
1024-Bit Input Register for Analysis
There are 1024 input bits in the custom ASIC that store and operate on user defined input data patterns. The INPUT_STRUCT_VALUE holding register routes to 32 different input data registers that are each 32 bits in length. The select for each bank of registers is cycled through from Bank0 to Bank31 as a total input vector is loaded into the custom ASIC. The input data registers are not externally readable or writeable by the user. The INPUT_DATA_WORD_COUNT register routes the INPUT_STRUCT_VALUE to the appropriate INPUT_REG_BANKn (see Table W below).
The user's access to input data patterns is through the INPUT_STRUCT_VALUE register. The user can indirectly load the input registers through writing a new input data pattern into the custom ASIC, by reading and modifying a value stored in DRAM, and by clearing a value stored in memory. The system also is responsible for loading these registers during the operation of generating new randomizer values for a new equation. Table X below describes operations that result in loading the 1024-bit input register.
1024-Bit Input Register for Classification
There are 1024 input classification bits in the custom ASIC that analyze received data patterns for parallel classification. The INPUT_CLASS_VALUE holding register routes to 32 different input data registers that are each 32 bits in length. The select for each bank of registers is cycled through from Bank0 to Bank31 as a total input classification pattern is loaded into the custom ASIC. The input classification registers are not externally readable or writeable by the user. The INPUT_CLASS_WORD_COUNT register routes the INPUT_CLASS_VALUE to the appropriate INPUT_CLASS_REG-BANKn (see Tables X and Y below).
The user's access to input classification patterns is through the INPUT_CLASS_VALUE register. The user directly loads the input classification registers through writing a new input data pattern into the custom ASIC
Masking and Enabling Logic and Masking Engine Details
The masking and enabling logic and the masking engine in the custom ASIC handle all aspects of the programmable masking function. The first major function is storage of user masking information including a description of the input data bits that are to be masked, as well as information identifying what bits are to be masked at what step of the sequential masking process. The second major function is the circuitry to generate the fixed masking of specified data bits. The third major function is the mask impact calculation circuitry. This circuitry drives the input data to determine the effect of each data bit on the output vector, so that when the bit is masked, it's effect can be cancelled out when interpreting a randomizer value. The fourth major function involves interpreting the outputs of the data framer to account for masking.
User Masking Information
The first level of masking information that must be stored by the system is the masking of all data bits beyond the length of data that the user desires to analyze, and the masking of data within a fixed OFF/ON period after the start of the input data pattern. To implement the length and OFF/ON period masking, a set of 1024 input enable bits is used to gate the input data bits to a zero in the cases where the bit is to be ignored. The INPUT_DATA_LENGTH register in the input section identifies the number of data bits that must be evaluated. The following registers identify the bit range that is to be masked by the OFF/ON period, and they are both user readable and writeable. The enable bits dictate whether the register are used and it's value is sent to the data framer (see Table Z below).
The user is supplied with four 32-bit input blocks that can be used for fixed and/or programmable bit masking. To achieve the effect of forced masking for bits covered by each enabled Programmable Mask register, the user must explicitly mask the bit pattern for each masking step pattern that is used. The 32 k bit input blocks occur on 32-bit word boundaries, and can be described with simple five bit numbers (1024/32=32 or five bits). These 5-bit values identify bits in the received data pattern that are captured by the data framer in it's masking registers. In addition, a sixth bit is used to identify whether the mask register is to be enabled or disabled. This impacts whether the data framer captures the corresponding data. Disabling unnecessary masking conserves overall system power consumption. All of the mask register enable bits are set to 0 at power-up so that masking is disabled. All of the mask register start registers are both user writeable and user readable (see Table M below).
The custom ASIC provides up to eight selective masking steps, and permits the user to use any combination of these steps for a specific data circumstance. Selective Mask Step 0 is always used as the initial masking step when a packet is received. Selective Mask Steps 1-7 are used to handle occasions where the user wishes to mask subsets of data based upon specific packet types or parameters. Each of the 32×4=128 maskable bits have eight selective masking bits associated with them for a total of 1024 selective masking step bits. All selective masking step registers are both user readable and user write-able. If a MASK_REGISTER is disabled, the selective step bits are ignored (see Table AB below).
In Table AB, the addresses 0x2A, 0x2B and 0x2C have been expanded for bits 0 to 4 for illustration purposes. In this example in selective masking pattern 2 for mask register 1, bits 1-4 are masked. In selective masking step 3 for mask register 1, bits 1 and 3 are masked. In selective masking step 4 for mask register 1, bits 0 and 2-4 are masked.
Forced Masking Circuitry
The custom ASIC includes 1024 bits that are used to enable or disable the bit from consideration. These bits are used to drive the selection of whether an input data bit drives the mapper circuitry, or whether a “0” value is switched in to mask out the value (see table AC below). All bits that are contained within the fields of enabled Programmable Mask Registers will automatically be enabled for purposes of driving the mapping circuitry.
Within the masking and enabling logic, there is a block of initialization circuitry that can load up the INPUT_ENAB_BANKn registers. The steps of this initialization include: setting all of the bits to enabled, clearing out all enables after the INPUT_DATA_LENGTH value, and disabling bits that occur between the MASK_OFF_CYCLE_REG and the MASK_ON_CYCLE_REG. The forced masking steps affect every input data pattern, and as such, it is expected that they are setup once upon initialization. If any of these registers are changed, the entire custom ASIC and respective mappings become invalid, and require recalculation. The forced masking state machine is started after the appropriate user setup sequence (see Table AD below).
Mask Impact Calculation Circuitry
The masking impact of each bit that is in one of the four 32-bit masking registers must be calculated for each equation that is used. As a new equation is swapped in, theses bits must be calculated prior to re-evaluating all of the available inputs. Each input bit can affect any of the sixteen output bits in the primary randomizer pattern. Therefore, there are a maximum of 16×4×32=2,048 bits that must be stored per each equation. With a set of eight equations, this translates to a total of 8×2048=16,384 masking impact bits. Because it is critical that the masking impact bits be available in real time for equation calculations, it is important that these mask impact registers be stored on the custom ASIC. The masking impact registers are internal, and are not user-readable (see Table AE below).
The calculation of the masking impact bits is done by injecting a single “1” pattern into every appropriate bit in the input data pattern. As an input of a “1” in a specific masked bit is applied, the output of the mapper circuit is a 16-bit mask impact that needs to be stored. This mask impact can then be calculated for each of the eight active equations before the walking “1”s pattern is advanced (see
The input source select directs whether the input register or the walking one's pattern should be used to drive the mapping circuitry (see Tables AF and AG below).
The walking ones pattern must be driven by the system into the inputs that are associated with each MASK_REGISTER. A state machine sequences through the registers to accomplish this task (see Table AH below).
The mapper outputs must be routed to the appropriate mask impact registers for the equation that is being addressed. For instance, when equation 1's mapping is being checked for MASK_REG0 and the system is looking at the impact of bit 31 in the register, the mapper value is written to the register EQ1_MASK_REG0_IMP_BIT31. The mapper outputs must be routed to a total of 1024=(8 equations*4 mask registers*32 bits ) mask impact registers. The novel custom ASIC must be able to generate an enable for each of these registers that is based on the equation number, the bit number, and the MASK_REGISTER.
Output Mask Adjustments
When a randomizer value from the data framer is received, the system must adjust it by canceling out any bits that have been masked. This operation is done using the mask impact registers and effectively adding the appropriate values into the received word based on the equation being used at the time. In the case of the primary and secondary randomizer values, different equations are used to generate these numbers (see Table AI below).
The storage registers for the mask capture data from the data framer are used in the calculation to cancel out the effect of masked bits (see Table AJ below).
Parallel Classification Output Mask Adjustments
When a parallel data pattern is loaded into the system from the microprocessor interface for classification, the system must adjust it by canceling out any bits that have been masked. This operation is done using the mask impact registers and effectively adding the appropriate values into the received word based on the equation being used at the time. In the case of the primary and secondary randomizer values, different equations are used to generate these numbers (see Table AK below).
The storage registers for the mask capture data from the data framer are used in the calculation to cancel out the effect of masked bits (see Table AL below).
Illustration of Programmable Masking
To determine the post masking result for any bit in a randomizer pattern, there are a total of four values that must be considered, i.e. the non-masked randomizer output (PRIM_RANDOMIZER_RX or SEC_RANDOMIZER_RX), the sequential mask step for the specific bit for the specific equation being analyzed (MASK_REGm_SMSp_EQn), the mask impact for the specific bit (EQn_MASK_REGm_IMP_BITq), and the actual captured status of the masked bit (MASK_CAPUTURE_DATA_m). From a masking perspective, there are a total of 128 (32×4) bits that can be masked. At this point, the mask registers, the mask capture data registers, and the mask impact registers are all relative, and we are not concerned with the absolute location within the data word. What is important is whether a specific bit is being masked in a specific selective masking sequence, whether the bit impacts the output vector, and whether the bit was actually captured as a one. Due to the nature of the randomizer implementations, data bits that are 0 do not impact the output vector (see Table AM below).
The programmable masking function can operate on either the primary or the secondary randomizer value. The system must be able to switch between these between any two cycles because evaluation of the primary and secondary values must be done in successive operations when searching the database. The RANDOMIZER_SELECT register determines which of these two randomizer values is analyzed.
The novel custom ASIC permits analysis of either data (serial) or parallel data passed over from the microprocessor. The SOURCE_SER_PAR signal is used to select between these two sources of data. This signal permits the system to operate while receiving packet classification information through these two possible interfaces (see Table AO below).
To illustrate the programmable masking function, consider one of the possible 128 bits that can be masked (see
In this example, an XOR tree masks the 128 possible masked bits with the selected randomizer output. The system permits a sequential selective masking approach (see Table AP below). In a sequential masking approach, the output of the first masking must point to a location in memory.
The system allows the user to specify a sequence of masking steps that can be taken for a given data pattern. In other words, the primary randomizer output can be changed for bits 0-4 of one of the masking words in the first pass, and then bits 5-7 in the second pass. The masking steps are provided in the primary randomizer table entries. The custom ASIC could be developed for expansion of the ability to logic analyzer capture serial data in the event a simple change to the data framer is made.
Processes Associated with the “Masking and Enabling Logic” Block
Table AQ below shows processes are used to setup the enable bits and to calculate the masking Impacts for each selected bit.
Equation Mapper
The equation mapper takes the latest pattern in the input register, as post operated on by the masking logic, and maps it through a giant XOR gate logic tree to derive a corresponding randomizer value. On average, each randomizer output bit is made up of half the input data bits. In this case, each output is the XOR of 512 inputs, An XOR tree for 512 inputs takes a total of (256+128+64+32+16+8+4+2+1=) 511 two input XOR gates. Each mapping uses 16 bits, or a total of (16*511=) 8176 gates to produce the randomizer output. When more and more equations are used, there is significant redundancy in the logic gates making up this XOR tree.
The analysis shown in Tables AR and AS below breaks down the tree into levels. The first level operates on groups of two inputs from the input data pattern. If these inputs are referred to as A and B, there is a single XOR term possible (A XOR B), and there are two single terms that flow through to the next level: A and B. At the second level of the tree, one can view the combinations of the groups (A,B) and (C,D). There are a total of six inputs for this block at the second level (A, B, A XOR B, C, D and C XOR D). When working with the tree, each of the inputs from the A,B group can be XOR'd with each of the inputs from the C,D up. This results in a total of nine possible XOR terms at this level. The output ms from each higher level=(Output Terms from previous level ˆˆ2+2*Output ms from previous level). The 2*Output terms from previous level accounts for pass through values.
Mappper Multiplexer
The mapper multiplexer circuitry must take equation mapper outputs and select the appropriate randomizer bit pattern for the equation that is being used. The size of the mapper multiplexer is dependent upon the number of equations being used. The register sown in Table AT below is used to signify the mapping that is to be selected.
The mapper multiplexer size can be calculated using the formula that a multiplexer tree for 2ˆn bits contains 2ˆn-1 2:1 multiplexers. Table AU below shows the number of multiplexers that are needed as a function of the number of equations that are implemented.
The output of the equation mulitplexer corresponds to the calculated randomizer value for the selected equation, and it is used to drive the mapper storage control block.
Mapper Storage Control and Storage State Machine
The mapper storage control and storage state machine handles storing and retrieving randomizer values and associated table information. When new inputs are provided, the system must calculate and store randomizer values for all equations of interest. When the system receives a randomizer value from the data framer, it must access these tables to identify the proper input.
Primary Randomizer Output Vectors
The output of the primary randomizer is a 16-bit number that maps to an input. For instance, if Input #5356 produces a value of 24593 in the primary randomizer, then whenever the primary randomizer value of 24593 is received, the system returns input #5356. Tables AW and AX below show how inputs map when randomized by different equations (A, B, C, . . . n).
In the above Table AX, there are six example locations with no inputs that map into a specific state: A map 2, B map 0, B map 3, n map 1, n map 2, and n map 65535. There are four example locations with one input that maps to a specific state: A map 0, A map 65535, B map 1, and n map 3. There are three locations with two inputs that map to a specific state: A map 1, B map 2, and B map 65535. There is one location where three inputs map into a specific state: A map 3. There is one location where four inputs map into a specific state: n map 0.
The example in the above Table AX has a much higher rate of one, two, three, or four inputs being mapped into a specific state than would be found in an actual implementation. This has been done for illustrative purposes only.
Primary Randomizer Table Entries
This discussion describes all of the entries in the primary randomizer table for the various possibilities of: no match, single match, pair match, triple match, quadruple match, overflow and masking (see Tables AY, AZ, and BA below).
The determination as to where to check the secondary randomization pattern is left to a later time. This is due to the fact that the user could chose to ignore the secondary randomization pattern. Of special interest is the case where a single match occurs that is a masking value and not a final input value. This case is handled in the Triple/Overflow/Mask case, where there are sufficient bits to encode the situation. In many ways, the MASK situation can be viewed as a multiple input situation (see Table BB below).
The Multiple Match Table Entry points to one of 1024 multiple match table entries (see Table BG below).
The Multiple Match Table Entry points to one of 1024 multiple match table entries (see Table BG below).
The Multiple Match Table Entry points to one of 1024 multiple match table entries (see Table BG below).
The multiple match table entry points into the multiple match table, with the exception that the table format changes so that only the input numbers are stored and not the secondary randomizer numbers. This allows up to eight inputs to be stored. Because an overflow table entry evolves from a quadruple table entry, one does not move the four inputs that were previously stored, but the new inputs overwrite the existing secondary randomizer values in the structure.
The number of overflows(#Over) value specifies how many inputs are stored in what is normally the secondary randomizer section of the multiple match table entry structure. If there is a single overflow, this value is ‘01’. If there are two overflows this value is ‘10’, if there are three overflows this value is ‘11’, and if there are four overflows this value is ‘00’ (see Table BF below).
Multiple Match Table
The custom ASIC uses a 1024 entry multiple match table (see Table BG below) to handle the cases where two, three, four, or the overflow case of (5-8) input vectors map to the same primary randomizer output. For each possible pair, there are two possible inputs and associated secondary randomizer values that need to be stored. For each possible triple, there are three possible inputs and associated secondary randomizer values that need to be stored. For each possible quadruple, there are four possible inputs and associated secondary randomizer values that need to be stored. In cases of 5-8 inputs which are signified as being overflow cases, the secondary randomizer values are dropped. The primary randomizer table entry dictates the inputs and secondary randomizer values in the multiple match table entry that are valid.
Multiple Match Table Valid Entries
The VALID_MULT_ARRAY (see Table BH below)stores information as to which multiple match table entries have been used by a particular mapping. There are 1024 multiple match table entries associated with each mapping, and the VALID_MULT_ARRAY is therefore 1024/16=64 values long to identify where there are open multiple entries. In addition, a second step of having four 16-bit values to signify where there are available entries that are used. One could conceivably use a pointer or pair of pointers in the ASIC to keep track of the next available inputs.
Storage Required for each Equation
Table BI below shows the memory storage required for each mapping that is stored.
Registers used in Storing and Reading a New Randomizer Entry
Table BJ below lists the registers used to access and modify the randomizer table. The output of the mapper is the CALC_RANDOMIZER_VALUE. The selection of the equation mapper uses the EQUATION_MAP_SELECT register.
The storage value is EQUATION_STORE_ENTRY which dictates where the value is stored in the primary randomizer table. In the case of storing a new value into the randomizer table, the INPUT_DATA_NUMBER is the output of a multiplexer that selects between the USER_INPUT_DATA_NUMBER and the SYS_INPUT_DATA_NUMBER and it contains the input number for the value while the PRESENT_MASK_STEP and NEXT_MASK_STEP store the necessary masking information.
When interpreting a randomizer value from the data framer, the RANDOMIZER_SELECT value is used to determine whether to analyze the primary or secondary randomizer values. The PRIM_RANDOMIZER_RX register contains the primary randomizer value, and the SEC_RANDOMIZER_RX register contains the secondary randomizer value. The PROG_MASK_RX signal is the masked value of the received primary or secondary randomizer signal (see Table BJ below).
Registers Associated with Initializing all Randomizer Mappings
The following registers are associated with clearing out a randomizer table and with clearing out the multiple table entries for any equation. Due to the fact that the system may be dynamically clearing out an equation entry in the middle of updating new equations, it is important that there be independent registers to point into the SRAM for the initialization process (see Table BK below).
Processes for Randomizer Table Manipulations
There are a number of independent operations to manipulate the primary randomizer tables. These involve initialization, storage and retrieval of data (see Table BK below).
Mapper Engine, Statistics and Equation State Machine
The purpose of the mapper engine, statistics, and equation state machine is to cycle through the various equation mappings for new inputs, maintain statistics for each equation mapping, select the appropriate equations to use, and initiate swapping out equations that do not produce appropriate mappings. This block has a state machine that operates on inputs across all equations. It implements counters in hardware to maintain statistics for each equation.
Primary Randomizer Equation Analysis
The system maintains data for a total window of eight primary randomizer equations on-chip. The system selects between these eight mappings to pick the best one using a variety of characteristics. Each of these equations have counters associated with the number of triples, quadruples, or overflow values that they contain. Based upon programmable criteria (values written into registers), the user can decide how often an equation is deemed to be non-usable and a search is initiated for a replacement. This directly affects power consumption because any time a search for a better equation is done, there are a large number of accesses to DRAM. On the flip side of the equation, when the equation set is static, there is extremely low power consumption from the custom ASIC.
Equation Status Counters
The following set of counters (Table BL below) is maintained for each equation on the custom ASIC. The output of these registers is used to determine the best equation for use, and whether certain equations need to be switched out.
The following error conditions are serious for a specific equation, and must be handled quickly by the custom ASIC:
The number of Overflow Entries (EQn_OVERFLOW) exceeds 0
The number of Multiple Entries (EQn_MULTS) exceeds 1024
In the worst case, a serious error results in an incoming packet not being identified by the custom ASIC. Any of the errors listed above would result in a vector not being found in the primary mapping table. A large majority of packets, i.e. 9999/10000, would still continue to be received as expected, and quickly, the chip would change the primary mapping equation to handle the serious fault.
Optimal Primary Equation Selection
The issues associated with picking the optimal primary equation are numerous. The greater the number of multiple entries, the longer the average lookup time is for the equation and the greater the probability that the multiple table entries is exceeded. The number of quadruples is an early indicator of possible problems because a quadruple is a single new input vector away from creating a serious error condition. Any equation that has reached an overflow condition for must be eliminated out of hand if possible. If no equation meets these criteria, then it is critical that the least offensive of the remaining equations be used. It is also critical that only equation mappings that contain all inputs be considered in this evaluation. This is necessary because there are times when equations are swapped out, and new ones are evaluated.
The following equation shows how the EQm_DISABLE bit is calculated. This bit is necessary to be able to shut down the use of equations that are inappropriate. The thresholds for quadruples and-triples have been set to very high levels that have an extremely low probability of occurrence in real life, and that pose a added burden to the receiver should they be used.
The equation optimization comparator input word for each of the eight possible equations has been designed so that the best equation is the one with the lowest word. This way, the eight equation optimization comparator input words can be sent through a 4→2→1 tree of 2:1 comparators to determine the optimal equation. In looking at the word, the highest priority for disqualification occurs when the equation table is incomplete. The second highest priority is that the equation has been disabled due to either overflow conditions or an excessive number of quadruple or triple matches. Next, the number of overflow conditions is used, which it is hoped is zero. The number of multiple inputs comes next in the priority structure, followed by the number of quadruples and triples in that order.
The equation optimization comparator produces the OPTIMAL_EQUATION output that signifies the best equation mapping at the present time (see Table BM).
The system identifies equations that are not usable, i.e. EQm_DISABLE==1, and determines when they reach a threshold programmed by the user (EQ_UPDATE_THRESH). When this threshold is reached, the system sets the EQm_INCOMPLETE bit for each disabled equation, clears out the disabled equations, and then updates the disabled equation (see Table BN below). This decision process also relies on equation aging which is described below.
Method to Track Mappings to Equation Values
There are eight different primary equation mappings that are used at any time by the custom ASIC. These equations map to one of 128 equations that are implemented in the mapper, which in turn map to one of the 32768 possible feedback paths for the randomizers. Actually, the important thing is that one stores a secondary mapping with each table for reference, and a way to determine whether there are duplicates. Storing whether a secondary mapping produced a duplicate would require a new bit. Primary and secondary randomizer values must be stored in the equation map for use because it is necessary to have their mask impact bits available on custom ASIC. Therefore, one needs to make sure that recently used primary and secondary randomizer equations are available.
Table BO below lists the registers associated with storing the mapping information. These values can be directly applied to the mapper multiplexer to determine the value. The system has 128 possible mappings, and uses a total of eight equations that each correspond to one of the mappings. Any time that an equation is found to be bad, a value of eight is added to the equation number. This guarantees that one will never have two equations that map to the same value. At any given time, one should use the best available mapping as the secondary randomizer equation value for new equations. The odds of this equation going bad is significantly less than other equations in the table, and the problem of being forced to keep it around has similarly less impact.
Maintenance of Equation Mappings
It is critical to maintain primary randomizer equation mappings that have been sent to the data framer for a reasonable time period. This is necessary to avoid a problem where an equation is swapped out and a value is read from the data framer that has no corresponding table for evaluation. In the case of parallel modes of operation, it is possible to use any of the equations and this is not a problem.
When a user is using the data framer, it is presumed that there is a time critical nature to the analysis of data packets. For this reason, it is possible to use a set of timers for each equation to indicate how long ago the equation was used in a data framer. These timers are a maximum of one second in length, and once one second has expired, the equation is considered as having been aged out. In the future, it may be valuable to permit a programmably variable shorter time to indicate that an equation has been aged out.
The equation aging registers are clocked at a 4 msec rate. When new primary and new secondary randomizer equations are written into the data framer, their equation aging registers are set to 0. The primary and secondary randomizer equations that are used for the data framer are be the ones used for parallel classification. All other equation aging registers are permitted to count upward. When an equation aging register counts up to a value of 255, it stops to signify that the counter has aged out (see Table BP below).
Processes used for Mapping Analysis
There are a number of processes associated with initializing and maintaining information regarding the equations used by the system (see Table BQ below).
Time Accelerator Block
The purpose of the time accelerator block is to advance a received randomizer value through “n” cycles of time in a single hardware cycle on the custom ASIC. The custom ASIC is running at extremely high speeds, and it is desirable to shut down the randomizers whenever possible to avoid power consumption. The calculated randomizer values in the custom ASIC are based upon a 1024-bit input word being used for the calculations, with all 1024 input bits being shifted into the randomizer. The custom ASIC has been structured such that bits occurring after the user selected data length are set to zero in the calculation (see
To solve this problem, the time accelerator block has been added to the custom ASIC. This block is able to take a randomizer value and shift it forward by the equivalent of “n” cycles of zero clocked input data all within a single cycle. The theory behind this time acceleration revolves around the fact that any shift of “n” cycles can be viewed as a remapping of the initial state of the randomizer stage values (qinit0, qinit1, . . . qinit15). To shift forward by a specific time of “n” cycles requires a specific remapping of the initial values to the final values. To accomplish a variable shift of any chosen value of “n” cycles, the novel custom ASIC implements a binary weighted programmable shifter. To accomplish a variable shift of from 1 to 1024 bits in length, the system implements a 512-bit shifter, a 256-bit shifter, a 128-bit shifter, a 64-bit shifter, a 32-bit shifter, a 16 bit-shifter, an 8-bit shifter, a 4-bit shifter, a 2-bit shifter, and a 1-bit shifter. Based on the selected shift value “n”, each of these fixed shifter stages is either switched into the data path or bypassed. Each shifter stage relies entirely on its own inputs to produce a direct mapping to its own outputs. There is no interaction between groups of stages other than the fact that the individual stages are producing a one-to-one mapping of inputs to outputs. To simplify this block, and the time associated with making calculations, the time accelerator may be modified to advance in 32-bit increments only. This would permit 512-bit shifts, 256-bit shifts, 128-bit shifts, 64-bit shifts, and 32-bit shifts only, but does nothing to affect the theory of operation for this block.
Time Accelerator Register The TIME_ACC_CYCLE is used to setup the number of cycles of acceleration to be applied to a received randomizer value. The source of the time accelerator block is chosen by the RANDOMIZER_SELECT value that chooses between the primary and secondary randomizer values that have been received. The equation that is being used in the analysis of the randomizer is critical for determining the mapping, and it is stored in the EQUATION_STORE_ENTRY register (see Table BR below).
Time Accelerator Logic Stage
The preferred implementation of one of the time accelerator stages using a large number of XOR gates and a smaller number of multiplexers (see
The example above shows how an individual bit in the output is calculated. This is multiplied by sixteen to handle each output bit in the remapping situation. The benefit of this approach is that the XOR trees for all of the equations, and for all of the output bits are shared. There is a limit to how many XOR gates can be used when there are a total of only sixteen inputs.
This approach uses a maximum of 428 XOR gates in it's implementation for the entire stage, and it uses only 16× 128:1 multiplexers which each contain 127 2:1 multiplexers. This is a total of 2,460 gates per timing accelerator stage.
Overall Timing Accelerator Architecture
The timing accelerator (See
Interface
The interface provides the functions of configuring the data framer and reading and interpreting data when a packet is received.
In configuring the data framer, the data length must be provided. The INPUT_DATA_LENGTH register described herein must be transferred to the data framer upon initialization
In configuring the data framer, a set of masking registers need to be initialized. The MASK_OFF_CYCLE_REG and MASK_ON_CYCLE_REG registers are described herein. In addition, MASK_REGISTER—0, MASK_REGISTER—1, MASK_REGISTER—2 and MASK_REGISTER—3 are described in the context of their start values. These six masking registers have associated enable bits to determine whether or not they must be loaded into the data framer. All masking registers are expected to be loaded into the data framer upon initialization or reset, but are not expected to be changed during operation.
In addition to the MASK and length registers, the data framer randomizer feedback registers must be configured. These are described in Table BT below, and their calculation is described later in this discussion.
In receive mode, the custom ASIC latches data from the data framer so that it can be analyzed. A FIFO structure may be necessary to permit packets to back up if necessary. With the new parallel mode of operation, it becomes more likely that an operation may preclude immediate access to the randomizer state machines because many channels could be using the same custom ASIC. Ideally, the interface always has the highest priority because it is the high performance interface. The implementation that is listed below supports storing a single register snapshot, but it could be easily increased to being a set of FIFOs.
In receive mode, the received randomizer values must be read from the data framer. The discussion herein describes the PRIM_RANDOMIZER_RX and the SEC_RANDOMIZER_RX registers that contain these two values. In addition, mask capture data must be received from the data framer, and this is stored in the MASK_CAPTURE_DATA—0, MASK_CAPTURE_DATA—1, MASK_CAPTURE_DATA—2 and MASK_CAPTURE_DATA—3 registers that are described herein.
Finally, the custom ASIC must know the randomizer feedback values that were used in calculating the primary and secondary randomizer values. If the feedback values have changed because the last packet, then a flag is set in the interface to show that these should be read. Otherwise, the custom ASIC can chose not to read these values. The description of these registers is found in Table BU below.
Equation Recovery Section
The feedback value that is returned from the data framer must be converted into a relative equation number from 0 to 7 that points to a primary randomizer table. To execute this feature, eight feedback register values must be stored along with their equation mapping (see Table BV below and
Randomizer Setup Section
Given the equation number, the proper EQn_FEEDBACK register can be selected to drive the data framer. This section shares the equation feedback registers with the equation recovery section. The OPTIMAL_EQUATION is driven in logic, and it is used to select the PRIM_RAND_FEEDBACK (see
Processes Associated with the Interface
The updating of the interface registers should be done as infrequently as possible to avoid churning. When a new equation is desired, new feedback registers must be latched into the data framer. On initialization, it is also necessary to load masking registers into the data framer.
Detailed State Diagrams
In all cases below, the process implementation tables are organized in the form: State name—activity—next state, whether or not status indicated in a Table heading.
State Machines for the “Input Manager Control and State Machines” Block
The following state machines are used to manage the user inputs in the system (see Tables BW-CN below).
State Machines for the “1024 Bit Input Register” Block
There are no state machines dedicated to this block.
State Machines for the “Masking and Enabling Logic” Block
The following state machines are used to manage the masking and enabling functions of the system (see Table CO-CR below).
State Machines for the “Equation Mapper” Block
There are no state machines specific to this block.
State Machines for the “Mapper Multiplexer” Block
There are no state machines specific to this block.
State Machines for the “Mapper Storage Control and Storage State Machine” Bik
The following processes have to do with initializing, storing, and removing and matching values in a primary randomizer table (see Tables CS-DV below).
State Machines for the “Mapper Engine, Statistics and Equation State Machine” Block
The following state machines are used to manage the mapper engine, statistic, and equation selection and calculation functions of the system (see Tables DW-EI below).
Process Summary
The following table lists a summary of the key processes described above.
Captured Packet Classification Description (Interface)
In
-
- The primary randomizer is latched into the FLAME_PRIM_RAND register.
- The secondary randomizer is latched into the FLAME_SEC_RAND register.
- The primary randomizer feedback value from the data framer ASIC is latched into the FLAME_PRIM_FB register.
- The secondary randomizer feedback value from the data framer ASIC is latched into the FLAME_SEC_FB register.
- Zero to four masking registers from the data framer ASIC are latched into the applicable FLAME_MASK_n registers.
A set of paths 190 are initiated when the master control block clears the system to start processing the received values. In the case of a single data framer ASIC, this is immediately.
-
- The path 190 to “2” shows the transfer of the FLAME_MASK_n registers into the MASK_CAPTURE_DATA_n registers. These transfers can be done entirely in parallel in a single cycle.
- The path 190 to “3” shows the time acceleration of the received randomizer values. The FLAME_PRIM_RAND value is time accelerated in a single cycle to generate the PRIM_RANDOMIZER_RX value. The FLAME_SEC_RAND value is time accelerated in a single cycle to generate the SEC_RANDOMIZER_RX.
Another set of paths 192 are initiated when the randomizer values need to be modified by masking prior to their use in accessing and verifying entries in the appropriate primary randomizer table.
-
- The path 192 from “2” to “3” shows how the MASK_CAPTURE_DATA, in conjunction with the mask register impact data for the appropriate equation and the selective masking step information is added together with the randomizer value.
- The path 192 from “3” to “4” shows how a properly masked randomizer value is directed to the mapper storage state machine.
- The path 192 from “4” to “2” reflects how when a new masking step is reached in the mapper storage state machine, the system must re-mask the randomizer values for their use in the next step of the mapper storage state machine.
Another set of paths 194 are used to show how the mapper storage control and storage state machine accesses the external SRAM.
A path 200 is initiated when a match in the primary randomizer table has been found.
-
- The path 200 from “4” to “6” shows how the matching input value is transferred to the microprocessor interface.
- In circumstances where no match is encountered, that information must be transferred to the microprocessor.
A path 198 is initiated when a value is ready for the microprocessor to read.
-
- This path generates an interrupt to the microprocessor.
- All necessary information regarding the match is contained in a register.
Captured Packet Classification Description (Parallel Microprocessor Interface)
In
-
- The system must write the parallel classification data into the INPUT_CLASS_REG_BANKn registers. This is done indirectly by writing into the INPUT_CLASS_VALUE register.
- When a write to the INPUT_CLASS_REG_BANKn is to a mask register that is being used, the system loads the parallel classification mask register.
One set of paths 216 are initiated when the master control block starts the parallel classification process.
-
- When the INPUT CLASS_WORD_COUNT exceeds the INPUT_DATA_LENGTH, the master controller is interrupted to start the parallel classification process.
- The path 216 from “1” to “3” shows how the INPUT_CLASS_REG_BANKn registers are masked with fixed enabling logic, and then mapped and latched into the PRIM_RANDOMIZER_PAR and SEC_RANDOMIZER_PAR registers using the optimal equation at the time.
- The path 216 from “2” to “3” shows how the appropriate EQn_MASK_REGm_IMP_BITx (mask impact bits) are applied based on the selective masking step being used to generate the masked randomizer value.
- The path 216 from “3” to “4” shows how a properly masked randomizer value is directed to the mapper storage state machine.
- The path 216 from “4” to “2” reflects how when a new masking step is reached in the mapper storage state machine, the system must re-mask the randomizer values for their use in the next step of the mapper storage state machine.
Another set of paths 218 are used to show how the mapper storage control and storage state machine accesses the external SRAM.
Another path 212 is initiated when a match in the primary randomizer table has been found.
-
- The path from “4” to “6” shows how the matching input value is transferred to the microprocessor interface.
- In circumstances where no match is encountered, that information must be transferred to the microprocessor.
Another path 210 is initiated when a value is ready for the microprocessor to read.
-
- This path generates an interrupt to the microprocessor.
- All necessary information regarding the match is contained in a register.
Probability Analysis
The following analysis is used to investigate the probability of any given primary and secondary mapping fitting within the constraints of multiple values mapping to the same endpoint. The basic problem involves the random distribution of mapped input values across an output space of a given size. As an example: if the output space has a size of 2ˆ16=65,536 values, and there are 10,000 random inputs, how many output values are mapped to by zero, one, two, three, four, or more inputs in a random probability analysis. This affects the chances that an individual mapping is usable, and directly impacts the number of mappings that should be maintained at any given time.
General Pairing (and higher) Odds Calculations
Let us use a problem where we have ten inputs that map to outputs with the letters A to J respectively. As we analyze the probabilities of various scenarios for these odds, we use a generic number of output vectors called “STATES” to signify the possible outputs to which any input can be randomly mapped. For purposes of probability analysis, a pair refers to two inputs mapping to a single output, a triple refers to three inputs mapping to a single output, and a quadruple refers to four inputs mapping to a single output.
Odds of No Pairs
In the case of having no paired outputs, each input must be compared to all of the others. The easiest way to analyze the output situation is to show what equalities do not occur as opposed to what equalities do occur. In the case of outputs being equal, they could be equal because they are components of a pair, a triple, a quadruple, or higher. In the cases of outputs being unequal, the situation is unambiguous. In the following tables, the “!=” sign is used to signify that two outputs are not equal.
When the first output “A” is analyzed, it must be compared against all other outputs. In the comparisons for “A”, the odds that “A” is not equal to any other output are (STATES-1)/STATES because there are STATES-1 non “A” outputs in a total of STATES outputs remaining. When the second output “B” is analyzed, the “A” state has already been eliminated. This reduces the possible remaining states to be used in the probability calculation since there are really only STATES-1 possible states, of which STATES-2 do not match “B” (see Table EK below).
Odds of Any Single Pair
When we are trying to calculate the odds of a single pair in this group of outputs, we can do the same the same sort of a table for a single possibility of a pair. When we are looking at a situation where A=B, the odds can be easily determined as 1/STATES. After the odds for a single pair are calculated, we must multiply by the number of possible pairs to determine the total odds of any pair occurring when ten inputs are mapped into “STATES” possible outputs (see Table EK below).
To determine the overall odds, the total number of possible pairs must be calculated. A binomial equation for probability provide us with the total number of pairs in ten outputs.
Binomial Equation:
In the case where n=2 and m=10, referred to as 2 choose 10, the binomial equation above produces a result of 45 possible combinations of a single pair in ten output values.
Odds of Any Single Triple
When we are trying to calculate the odds of a single triple, we are looking at a situation where A=B=C. The odds can for a single occurrence can be easily determined as (1/STATES)*(1/STATES). After the odds for a single triple are calculated, we must multiply by the number of possible triples to determine the total odds of a any triple occurring when 10 inputs are mapped into “STATES” possible outputs (see Table EL below).
The binomial equation for n=3, m=10 provides a total of 120 possible triples given ten possible output values. The odds for the single triple must be multiplied by 120 to get the overall odds for all cases of a single triple.
Odds of Any Single Quadruple
When we are trying to calculate the odds of a single quadruple, we are looking at a situation where A=B=C=D. The odds can for a single occurrence can be easily determined as (1/STATES)*(1/STATES)*(1/STATES). After the odds for a single quadruple are calculated, we must multiply by the number of possible quadruples to determine the total odds of any quadruple occurring when ten inputs are mapped into “STATES” possible outputs (see Table EM below).
The binomial equation for n=4, m=10 provides a total of 210 possible quadruples given ten possible output values. The odds for the single quadruple must be multiplied by 210 to get the overall odds for all cases of a single quadruple.
Odds of Pair and a Triple:
When we are trying to calculate the odds of a single pair where A=B, and a single triple where C=D=E, the odds are more complex. The odds for the single pair are (1/STATES), and the odds for the single triple are (1/STATES)*(1/STATES). All of the remaining inputs must not pair with either the output from the pair (A) or the output from the triple (C) (see Table EN below).
The calculation of the number of possible occurrences of a single pair and a single triple are slightly harder. For each possible pair, there are eight remaining outputs that can generate a triple. Similarly, for each possible triple, there are seven possible outputs remaining that can generate a pair. Approaching this problem from either side will produce the same answer.
Binomial Equation:
Total Possibility Calculation for a Single Pair and a Single Triple:
General Formula Development
By reviewing the above tables for “No Pairs”, “Single Pair”, “Single Triple”, “Single quadruple”, and “Single Pair and Single Triple”, it is possible to develop a generic formula to calculate the probability. The first term in the formula accounts for the pairs, triples and quadruples. This term accounts for the (1/STATES) terms for all of the possible outputs that are paired.
Once the pairs, triples, and quadruples have been accounted for, all of the appropriate outputs must be checked to make sure they do not match any other output. One element from each pair, triple and quadruple must be checked against one element from all remaining pairs, triples, and quadruples in addition to the remaining un-matched outputs. All remaining un-matched outputs must also be checked against each other to verify that they do not match. The UNMATCHED_OUTPUTS-1 term accounts for the fact that with n outputs, there are only n-1 checks to be done with the first one.
Calculation of Number of Occurrences
The number of occurrences (combo_hits) for any given combination of pairs, triples, and quadruples is calculated in a sequential form. The problem can be broken down first into calculating the combinations of quadruples within the space of inputs, then by subtracting out the inputs associated with quadruples and calculating the combinations of triples within the remaining space of inputs, and finally by subtracting out the inputs associated with quadruples and triples, and calculating the combinations of pairs within the remaining space of inputs. The combination values for pairs, triples and quadruples can be multiplied together to determine the overall odds for a specific combination of these elements.
The following examples show empirically how this is done, and show the derivation of the formulas for pairs and triples.
Explanation of combo_hits pair calculation for 10 Inputs
The number of possibilities is based purely upon the number of inputs that are present. If there are ten inputs, there are (10 choose 2)=45 possible pairs in ten inputs. Once the first pair is gone, there are (8 choose 2)=28 chances of selecting a second pair from the remaining inputs. Once the second pair is gone, there are (6 choose 2)=15 chances of selecting a third pair from the remaining inputs.
Given 45 possible pairs, there are (45 choose 3)=14190 combinations of three pairs out of 45 possible pairs.
The odds for the first pair are 45/45, the odds for the second pair to be non-overlapping are 28/44, and the odds for the third pair to be non-overlapping are 15/43.
Pair Example for 10 Inputs
If there are enough inputs left to make a pair:
Possible_Pairs=N_Choose—M(inputs-3*(triple_count)−2*(pair_count-1),2) Check remaining Inputs, How many possible pairs can be found. #1
Possible_Pairs/=(Total_pairs−(pair_count-1)) #2
Running_Pair_Hits*=Possible_Pairs #3
Combo_Hits=Running_Pair_Hits*N_Choose—M(Total_Pairs, Pair_count) #4
Test for five pairs:
{(10 Choose 2)/Total_Pairs}*{(8 Choose 2)/(Total_Pairs-1)}*{(6 Choose 2)/(Total_Pairs-2)}*{(4 Choose 2)/(Total_Pairs-3)}*{(2 Choose 2)/(Total_Pairs-4)}
Triple Example for Fifteen Inputs
If there are enough inputs left to make a triple:
Possible_Triples=N_Choose—M(inputs-3*(triple_count-1),3) Check remaining Inputs, How many possible triples can be found. #1
Possible_Triples/=(Total_triples−(triple_count-1)) #2
Running_Triple_Hits*=Possible_Triples #3
Combo_Hits=Running_Triple_Hits*N_Choose—M(Total_Triples, Triple_Count) #4
Total_triples=N_Choose—M(15,3)=455
Running_Triple_Hits=1
Test for five triples:
{(15 Choose 3)/Total_Trips}*{(12 Choose 3)/(Total_Trips-1)}*{(9 Choose 3)/(Total_Trips-2)}*{(6 Choose 3)/(Total_Trips-3)}*{(3 Choose 3)/(Total_Trips-4)}
Quadruple Example for 20 Inputs:
If there are enough inputs left to make a quadruple:
Possible_quadruples=N_Choose—M(inputs-4*(quadruple_count-1),4) Check remaining Inputs, How many possible quadruples can be found. #1
Possible_quadruples/=(Total_quadruples−(quadruple_count-1)) #2
Running_quadruple_Hits*=Possible_quadruples #3
Combo_Hits=Running_quadruple_Hits*N_Choose—M(Total_quadruples, quadruple_count)
Test for five quadruples:
{(20 Choose 4)/Total_Quads}*{(16 Choose 4)/(Total_Quads-1)}*{(12 Choose 4)/(Total_Quads-2)}*{(8 Choose 4)/(Total_Quads-3)}*{(4 Choose 4)/(Total_Quads-4)}
Multiple Table Analysis
Introduction
The following analysis is used to evaluate the odds that various combinations of multiple tables will be sufficient for handling all of the pairs, triples and quadruples that may arise. The issue is what the odds are that any given equation will be handled properly by the tables that are available. If this overall odds for any random equation is too low, then it will be necessary to store mappings for more equations and to be able to swap equations more frequently.
Primary Randomizer Simulations for Various Scenarios
The following Primary Randomizer simulations were run for 10,000 inputs and 65,536 states. The goal of these simulations was to show what probability of success could be achieved by permitting a certain number of pairs, triples and quadruples of Primary Randomizer values.
This experiment shows that allowing more than 48 Triples seems to produce minimal impact.
Use 48 Triples, and Vary Pairs with No Quadruples
This experiment shows that allowing more than 768 Pairs seems to produce minimal impact.
Use 896 Pairs, 48 Triples, and Vary Quadruples
This experiment shows that allowing more than 5 Quadruples seems to produce minimal impact.
Primary Randomizer Simulation Conclusions
The above experiments show that for the example of 65,536 States and 10,000 inputs, there is a better than 95% chance that a Primary Randomizer equation value will produce no more than 896 Pairs, 48 Triples, 5 Quadruples. By permitting 1024 Multiple Entries, that can be pairs, triples or quadruples, the odds of a successful Primary Randomizer will therefore be better than 95%.
Secondary Randomizer Probability Analysis
Introduction
Once the Primary Randomizer spreads out the inputs across the memory space, the job of the Secondary Randomizer is to differentiate between any paired, tripled or quadrupled inputs. The probability analysis for this function is much different than for the Primary Randomizer. For ease of analysis, worst case Primary Randomizer distributions can be used instead of weighting the values for all possible situations.
Pairs
In the case of a Primary Randomizer Pair, the odds that the uncorrelated Secondary Randomizer values will be the same can be calculated as follows:
In the case of a Pair, this is a straightforward probability where there are (STATES-1) out of (STATES) values for the second Secondary Randomizer Value that will not be a match.
Triples
In the case of a Triple with three Secondary Randomizer Values labeled A, B, and C, there are a number of odds that must be included. There are three individual possibilities that must be considered where A=B, B=C or A=C. Any of these could individually destroy the usability of the Triple, and they include the odds of all three being the same.
Quadruples
In the case of a Quadruple, with four Secondary Randomizer Values labeled A, B, C and D, there are a number of odds that must be included. There are six possible cases of a pair: A=B, A=C, A=D, B=C, B=D, and C=D.
Overall Secondary Randomizer Odds
In a case with a number of Pairs, Triples and Quadruples, the following formula can be used to determine the Overall Odds that the Secondary Randomizer will differentiate all Pairs, Triples and Quadruples.
Odds_Good_SR=Odds_Pair_Has_No_MatchNumber
Odds_Triple_Has_No_MatchNumber
Odds_Quadruple_Has_No_MatchNumber
Secondary Randomizer Example
For an example case of 65536 States, with 896 Pairs, 48 Triples, and 5 Quadruples, the odds that the Secondary Randomizer values will not be duplicated for any of the pairs, triples or quadruples can be calculated as follows:
In this example, there is a greater than 98%chance that the Secondary Randomizer will differentiate all of the pairs, triples, and quadruples. It should be noted that the odds that 896 pairs, 48 triples and 5 quadruples will be needed are very slim. In the average case, the odds that the Secondary Randomizer will be usable will be much improved.
Feedback Shift Register Theory
The purpose of this discussion is to explain the basics of serial shift register theory, and how those apply to the system. The areas that are analyzed include basic feedback shift register operation, future state prediction, shift register time acceleration, and masking of input values. A four bit feedback shift register is used to explain the theory behind the system (see
Basic feedback Shift Register Theory
An exemplary basic feedback shift register contains four D-flip flops that are clocked simultaneously. Each of these flip flops is also known as a stage in the shift register. With four stages, the shift register has 2ˆ4=16 possible states. In the case of the shift register in this example, the Q2 and Q3 stages are fed back to generate the first stage Q0 in conjunction with the INPUT value. This feedback mechanism occurs continuously for each clock of the CLOCK signal, and results in the values of Q0-Q3 cycling through a pattern that depends upon their initial state as well as the pattern of INPUTs that are applied.
Equations For Each Shift Register Stage
The next value of each shift register stage, i.e. Q0+, Q1+, Q2+ and Q3+, can be calculated as a function of both the present values of all shift register stages, i.e. Q0, Q1, Q2 and Q3, and the value of the INPUT.
Shift Register Equations
Q0+=((INPUT)XOR(Q2 XOR Q3))
Q1+=Q0
Q2+=Q1
Q3+=Q2
Features of XOR Gates and XORTrees
A 2-input XOR gate produces an output of “1” when it's inputs are different. If a 2-input XOR gate's inputs are the same, it produces and output of “0”. When more than two inputs are XOR'd together, the output is a “1” if there are an odd number of inputs that are “1”s and the output is a “0” if there are zero or an even number of inputs that are “1” (see Table ER below).
From the XOR logic truth table, it can be seen that the for an value “A”, the XOR of “A” and “A” equals “0”. This results because whether “A” is a 0 or a 1, when “A” is XOR'D with itself, the output is a zero. This feature is critical when evaluating shift registers over time.
Shift Register State Prediction
As a shift register is clocked along, the feedback continues to introduce values back into the input stage. In some cases, these cancel out terms that exist due to the fact that “A” XOR “A”=0. In other cases, terms continue to propagate. For purposes of description, the sequence of inputs that are applied to the shift register are A,B,C,D,E,F,G,H,J,K,L,M,N. In this example, the initial state is assumed to be Q0=Q0S, Q1=Q1S, Q2=Q2S, and Q3=Q3S to signify that these are starting values. For description purposes, the XOR of multiple inputs are XO)R(a,b,c, . . . z). In Table ES shown below, terms that are canceled out for a specific stage after a specific input are not shown.
From a theoretical standpoint, the exemplary shift register could be used to map inputs A,B,C,D,E,F,G,H,J,K,L,M, and N to a 4 bit value consisting of Q0, Q1, Q2 and Q3. Theoretically, after a shift register has been clocked through some number of cycles (thirteen in this case), the output can be predicted exactly by knowing what the inputs were and what the initial condition was. In the case of the system, the randomizers in the data framer ASIC are initialized to all 0's prior to the start of the packet. This removes the Q0S, Q1S, Q2S, and Q3S terms from the analysis since they are 0i. By doing this, the state of each stage becomes an XOR tree of a set of inputs.
The system could allow the randomizer to start at any value at the start of the packet. In that case, the value of the randomizer at the start of the packet would have to be captured and factored out of the final result. The mapped effects of the initial condition would have to be subtracted out of the final randomizer numbers knowing what the initial conditions were. This approach is similar to the masking approaches used in the system.
In the case of Table ES, Q0 after thirteen cycles is the XOR of inputs B,C,D,E,G,J,K and N. This XOR function can be implemented in a tree of XOR gates to generate the value. The values of Q1, Q2, and Q3 can be calculated using similar trees of XOR gates that use the inputs found in their specific equation.
By viewing Table ES, it can be seen that the XOR terms for a specific stage such as Q0 vary from clock cycle to clock cycle. Instead of providing length variability by attempting to calculate these trees for each length (which would be extremely prohibitive in size), the system uses a fixed length of inputs in its equation generators regardless of the end user length selection. The system relies on the fact that a “0” value input does not appear in the output state vectors, other than in the shifting of the state register that it introduces.
Time Acceleration Theory
To produce a general purpose ASIC, the system implements a large number of possible input values (1024). This large fixed value allows a single set of XOR trees to be implemented to calculate the state values for each equation. The result is that in most situations, the input values that the user is classifying are padded with trailing “0” values. In many applications, users may wish to classify only 50 to 100 possible inputs. In the data framer ASIC, the randomizers could continue to cycle “0” values into the randomizers, after the bits of interest, to reach a final value of 1024 that is used to generate all equation values. This results in added time delay, and additional power consumption in the high speed data framer. As an alternative, a feature called time acceleration is used.
If a feedback shift register, is applied a sequence of “0”s prior to being disabled, the final output state is a predictable value based on the state prior to the start of the shifting in of “0”s, and the number of “0”s that are shifted in. In Table ES, this can be viewed as having an initial state and then having all “0” inputs. In this case, the output vector after some number of cycles is purely a functional re-mapping through an XOR tree of the initial state. To provide flexibility, a set of binary weighted shifters are implemented. This permits from 1 to 1023 trailing zeros to be effectively simulated by the time accelerator.
As an example of how the successive binary weighting would work, Table ES can be used to show how values are mapped to generate different weightings for shifting 0's. In the case of Table ET below, each shift is determined by looking at the table above and removing all of the Input values because these are each 0. The values listed Table ET below reflect how a starting value of Q0S, Q1S, Q2S, Q3S is mapped after “n” cycles of “0” inputs.
The time acceleration approach involves selectively and serially mapping the inputs through a number of binary weighted “0” input shift stages to generate an output. For example purposes, a starting vector of Q0S, Q1S, Q2S, and Q3S are first mapped through a “0” input shift length of one. The result is then mapped through a “0” input shift length of two. That result is then mapped through a “0” input shift length of four. The result after mapping through shift lengths of 1, 2, and 4 should then equal a result as though we started with Table ES above and had seven cycles of inputs that were equal to “0” (see Table EU below).
When comparing with Table ES above, for the INPUT G row, and zeroing out inputs A-G, we can see that the mapping is equivalent to the table immediately above after a total shift of 1, 2 and 4 zero inputs. This displays how a variable time accelerated value can be rapidly calculated for a given equation.
Masking Theory
Referring to Table ES above, it can be seen that every input bit appears in an XOR tree term for one or more of the shift register output bits. As described earlier, if an input is XOR'd with itself, the result is a value of “0”. Therefore, if a final shift register output is known that included a specific input in its calculations, that input can be removed by XOR'ing each bit in the result with that specific input bit, if that input bit affected the final shift register output (see Table EV below).
In this case, if we wish to selectively mask out INPUT E from the result, we would XOR input E in with the Q0 and the Q2 bits. The selective masking approach in the system relies on the theory of knowing and subtracting out the effects of specific inputs. For any given Input bit that is desired to be masked out, it is possible to determine which bits in the output result are affected by the input bit. This is done by using the input mapper, for a specific equation, and sets the initial state of the randomizer to all “0”s, while controlling all inputs in parallel simultaneously. The process is done through a walking ones pattern that takes the bit under investigation and sets it to a “1”, while setting all other bits to a “0”. As each bit that is analyzed is set to a 1, the mapper outputs are stored in bit locations referred to as mask impact bits. The system allows the user to decide selectively what bits must be masked, versus those that can be always analyzed, or those that are never viewed.
Explanation of Masking Effect
The effect of masking is to remove an input from consideration in the analysis. This removal effectively makes the input appear as though it was a 0 and had no affect on the final shift register output. Therefore, there is no difference between a masked input and one that sets all of the masked bits to zero. In the following example of three inputs and their various combinations, input B is masked. The like shaded rows are equivalent as a result of this masking (see Table EW below).
Sequential Masking
A two step masking operation could be done as follows: first inputs A and B are masked and we differentiate based on input C. If input C is found to be a 0, then input B is masked and we differentiate on input A. If input C is found to be a 1, then input A is masked and we differentiate on input B (see Table EX-EZ below).
The example above shows how the results of an initial check can be used to drive what masking is done at the next step in the process. In the case that is illustrated, input C is first used to determine whether input A or B should be analyzed.
Although the invention is described herein with reference to the preferred embodiment, one skilled in the art will readily appreciate that other applications may be substituted for those set forth herein without departing from the spirit and scope of the present invention. Accordingly, the invention should only be limited by the Claims included below.
Claims
1. In a network for high speed transmission of digital data, said network comprising a memory, an apparatus for rapid differentiation between input data, comprising:
- a module comprising functional elements for each of adaptive, programmable, predictive, and sequential randomization of said digital data;
- said module comprising at least one element for receiving parallel input data, at least one randomizer that is driven by said input data, wherein a final state of said at least one randomizer is used as an index into said memory to determine which if any input data pattern has been matched;
- wherein input data pattern matching effects data classification.
2. The apparatus of claim 1, wherein general probability theory is used to evaluate randomization of said input data.
3. The apparatus of claim 1, wherein said element for predictive randomization comprises means for pre-calculating said randomization for each possible input data pattern.
4. The apparatus of claim 3, wherein said pre-calculation is performed at a time that a new input data pattern is entered.
5. The apparatus of claim 1, further comprising:
- means for performing a full hardware calculation of expected randomizer outputs for each input that is applied.
6. The apparatus of claim 5, wherein a plurality of randomizer feedback values are evaluated in real-time when a new input is applied.
7. The apparatus of claim 6, further comprising:
- means for storing said randomizer output values in said memory for each randomizer mapping that is being considered at a time.
8. The apparatus of claim 1, wherein said element for high speed prediction permits a significant number of possible randomization mappings to be maintained in said memory at any time.
9. The apparatus of claim 1, further comprising:
- means for adjusting a possible randomization mapping after any data have been received if an existing randomization mapping is significantly less ideal than another randomization mapping that has been evaluated.
10. The apparatus of claim 1, further comprising:
- means for maintaining statistics on substantially all presently evaluated randomization mappings to determine a best randomization, as well as any randomization that may be no longer usable.
11. The apparatus of claim 1, further comprising:
- means for quickly bringing substantially all of said input data patterns back to evaluate other possible randomization patterns when a randomization is no longer usable.
12. The apparatus of claim 1, further comprising:
- means for implementing sequential masking operations on said input data to accomplish sequential randomization of said input data patterns.
13. The apparatus of claim 12, further comprising:
- means for permitting either of fixed or programmable masking of selected bit patterns within said input data.
14. The apparatus of claim 13, wherein said masking operations permit a user to pre-program a series of masking decisions that can result in a final input data pattern match.
15. The apparatus of claim 1, wherein said element for adaptive randomization comprises means for adjusting said randomization, over time, to handle changing input data patterns that are to be analyzed.
16. An apparatus for rapid differentiation between received input data comprised of a limited number of input data bits, comprising:
- a randomizer for providing a usable randomization pattern, for a random set of parallel inputs, based upon an effective mapping of received input data patterns to output vectors; and
- means for handling a limited number of cases where two or more received input data patterns are mapped to a same output value.
17. The apparatus of claim 16, said means for handling further comprising:
- means for permitting a set value of multiple output cases where any of two, three, or four input data patterns map to a same output pattern;
- wherein said means for permitting evaluates a number of paired, tripled, and quadrupled output vectors in determining which randomizer mapping to use, as well as to determine when a randomizer mapping should be discarded.
18. The apparatus of claim 16, said randomizer further comprising:
- a primary randomizer for mapping each received input data pattern to an output value;
- wherein sufficient randomizer mappings are simultaneously evaluated to provide that a usable mapping is substantially always available.
19. The apparatus of claim 18, said randomizer further comprising:
- a secondary randomizer for differentiating between received input data patterns that have been mapped to a same output value;
- wherein entries in each of multiple data input patterns are different from each other.
20. The apparatus of claim 16, wherein for a given number of output states, a given number of input data patterns, and a given number of multiple outputs, said means for handling determines a probability that any specific randomizer maps said received input data patterns into a usable set of output states.
21. A method for ultra-high speed data classification, comprising the steps of:
- providing a data framer for framing input data;
- providing a complex circuit for permitting a user to differentiate between a plurality of different patterns in said input data; and
- performing parallel mode classification by providing fast classification of data that have already been stored in a memory as successive parallel words of data by performing adaptive programmable randomization.
22. An apparatus for ultra-high speed data classification, comprising:
- a data framer outputting parallel data;
- an adaptive programmable randomizer; and
- a complex circuit for controlling said adaptive programmable randomizer.
23. The apparatus of claim 22, wherein said complex circuit maintains multiple input pattern mappings associated with different primary and secondary randomizer equations, determines a best randomizer selection, decides when to switch randomizer values, and determines when a randomizer value is no longer useful and an entirely new mapping should be generated.
24. The apparatus of claim 22, wherein said complex circuit comprises any of the following:
- a microprocessor interface for communicating to a host processor system;
- a first memory interface for communicating with either of a stand alone memory or shared dual port memory, for storing data patterns to be matched;
- a second memory interface for communicating with a dedicated memory that contains mappings for a plurality of primary and secondary randomizer settings; and
- an interface for communicating with said data framer.
25. A method for ultra-high speed data classification, comprising the steps of:
- providing a data framer for framing received input data in parallel;
- providing a complex circuit for permitting a user to differentiate between a plurality of different patterns in said input data;
- said user loading input patterns into said complex circuit, said input patterns made up of a combination of data and, optionally, one or more masking steps that are performed on said input data in a potentially sequential fashion, said masking steps optionally comprising mapping a range of data values to a single mask step output and, when that mask step output is reached, executing additional programmable masking or verification;
- wherein said optional one or more masking steps are responsible for enabling only those input register bits that a user is interested in examining; and
- providing an equation mapper for generating a related randomizer value; said equation mapper performing the step of: calculating randomizer mappings for different equations simultaneously, wherein a randomizer value is calculated in a single cycle;
- wherein said complex circuit adjusts adaptively to select optimal randomizer settings based on input patters and masking patterns that have been applied to said complex circuit.
26. The method of claim 25, wherein input patterns are handled by an input manager control and state machine function where they are directed into an input register.
27. The method of claim 25, wherein said input patterns are loaded into an external memory for use in cases where a mapping is discarded and a new mapping must be generated.
28. The method of claim 25, further comprising the step of:
- providing a mapper multiplexer for immediate selection between each of a plurality of possible randomizer outputs associated with each of a plurality of possible randomizer equations.
29. The method of claim 28, further comprising the step of:
- providing a mapper storage control and storage state machine for saving and retrieving values from a plurality of equation mapping tables.
30. The method of claim 29, wherein said mapper storage control and storage state machine determines whether a present location pointed to by a primary randomizer value contains 0, 1, 2, 3, or 4 entries.
31. The method of claim 29, wherein said mapper storage control and storage state machine is responsible for handling creation and destruction of multiple entries, and adjustment of multiple entry tables that dictate those entries that are used.
32. The method of claim 25, further comprising the step of:
- providing a masking engine for permitting a user to setup sequential masking operations.
33. The method of claim 25, further comprising the step of:
- providing a time accelerator for re-mapping a received randomizer value to generate a randomizer value that would have been received if zero values had been clocked into a randomizer for a fixed number of cycles after a received randomizer value was captured.
34. The method of claim 25, further comprising the step of:
- providing a mapper engine, statistics, and state machine for determining equations to be used, and for determining when said equations are no longer usable and need to be replaced.
35. The method of claim 34, wherein said mapper engine, statistics, and state machine maintain statistics on all equation mappings that are maintained in a memory, selects a best mapping, and sends said to said data framer.
36. A method for ultra-high speed data classification, comprising the steps of:
- providing a data framer for framing input data;
- providing a complex circuit for permitting a user to differentiate between a plurality of different patterns in said input data;
- said user loading input patterns into said complex circuit, said input patterns made up of a combination of data and a plurality of masking steps that are performed on said input data in a potentially sequential fashion;
- wherein said input patterns are loaded into a memory for use in cases where a mapping is discarded and a new mapping must be generated;
- providing an equation mapper for generating a randomized value, said equation mapper performing the step of:
- calculating randomizer values;
- wherein said complex circuit adjusts adaptively to select optimal randomizer settings based on input patterns and masking patterns that have been applied to said complex circuit.
37. The method of claim 36, wherein input patterns are handled by an input manager control and state machine function where they are directed into an input register.
38. The method of claim 36, further comprising the step of:
- providing a mapper multiplexer for immediate selection between each of a plurality of possible randomizer outputs associated with each of a plurality of possible randomizer equations.
39. The method of claim 36, further comprising the step of:
- providing a mapper storage control and storage state machine for saving and retrieving values from a plurality of equation mapping tables.
40. The method of claim 39, wherein said mapper storage control and storage state machine determines whether a present location pointed to by a primary randomizer value contains 0, 1, 2, 3, or 4 entries.
41. The method of claim 39, wherein said mapper storage control and storage state machine is responsible for handling creation and destruction of multiple entries, and adjustment of multiple entry tables that dictate those entries that are used.
42. The method of claim 36, further comprising the step of:
- providing a masking engine for permitting a user to setup sequential masking operations.
43. The method of claim 36, further comprising the step of:
- providing a time accelerator for re-mapping a received randomizer value to generate a randomizer value that would have been received if zero values had been clocked into a randomizer for a fixed number of cycles after a received randomizer value was captured.
44. The method of claim 36, further comprising the step of:
- providing a mapper engine, statistics, and state machine for determining equations to be used, and for determining when said equations are no longer usable and need to be replaced.
45. The method of claim 44, wherein said mapper engine, statistics, and state machine maintains statistics on all equation mappings that are maintained in a memory, selects a best mapping, and sends said to said data framer.
Type: Application
Filed: Dec 14, 2006
Publication Date: May 24, 2007
Inventor: Brian Messenger (Monte Sereno, CA)
Application Number: 11/610,705
International Classification: H04L 12/56 (20060101); H04L 12/28 (20060101);
