Abstract: A system and method for storing error correction check words in computer memory modules. Check bits stored in physically adjacent locations within a dynamic random access memory (DRAM) chip are assigned to different check words. By assigning check bits to check words in this manner, multi-bit soft errors resulting from errors in two or more check bits stored in physically adjacent memory locations will appear as single-bit errors to an error correction subsystem. Similarly, the likelihood of multi-bit errors occurring in the same check word may be reduced.

Abstract: A multiprocessor computer system is configured to selectively transmit address transactions using either a broadcast mode or a point-to-point mode. Depending on the mode of transmission selected, either a directory-based coherency protocol or a broadcast snooping coherency protocol is implemented to maintain coherency within the system. A computing node is formed by a group of clients which share a common address and data network. The address network is configured to determine whether a particular transaction is to be conveyed in broadcast mode or point-to-point mode. In one embodiment, the address network includes a mode table with entries which are configurable to indicate transmission modes corresponding to different regions of the address space within the node. Upon receiving a coherence request transaction, the address network may then access the table in order to determine the transmission mode, broadcast or point-to-point, which corresponds to the received transaction.

Abstract: A pending tag system and method to maintain data coherence in a processing node during pending transactions in a transaction pipeline. A pending tag storage unit may be coupled to a cache controller and configured to store pending tags each indicative of a coherence state for a data line corresponding to a pending transaction within the transaction pipeline. The pending tag storage unit includes a total amount of storage which is substantially less than an amount required to store tags contained in the full tag array for the cache memory. When a pending tag exists in the pending tag storage unit, the coherence state of the corresponding data line within the cache memory is dictated by the pending tag for snoop operations. Accordingly, data coherence is maintained during the period when transactions are pending, e.g., not yet presented to a processor and cache.

Abstract: A probabilistic queue lock divides requesters for a lock into at least three sets. In one embodiment, the requesters are divided into the owner of the lock, the first waiting contender, and the other waiting contenders. The first waiting contender is made probabilistically more likely to obtain the lock by having it spin faster than the other waiting contenders. Because the other waiting contenders spin more slowly, the first waiting contender is more likely to be able to observe the free lock and acquire it before the other waiting contenders notice that it is free. The first of the other waiting contenders that determines that the previous first waiting contender has acquired the lock is promoted to be the new first waiting contender and begins spinning fast. Because only the first waiting contender is spinning fast on the lock, it is probable that only the first waiting contender will attempt to acquire the lock when it becomes available.

Abstract: A system for N-bit part failure detection using n-bit error detecting codes where n is less than N is disclosed. In a computer system having storage devices N bits wide and an error detection and correction capability of less than N bits, bit assignments are made so that storage device failures will be detectable because of the manner the effect of a part failure is distributed among multiple codewords. Consequently 8 and 16 bit wide DRAMs may be used in a memory system using error detection and correction codes which are not capable of detecting 8 or 16 bit errors in a codeword, and still preserve the ability to detect the worst errors possibly caused by a part failure.

Abstract: Low-latency distributed round-robin arbitration is used to grant requests for access to a shared resource such as a computer system bus. A plurality of circuit board cards that each include two devices such as CPUs, I/O units, and ram and an address controller plugs into an Address Bus in the bus system. Each address controller contains logic implementing the arbitration mechanism with a two-level hierarchy: a single top arbitrator and preferably four leaf arbitrators. Each address controller is coupled to two devices and the logical "OR" of their arbitration request is coupled via an Arbitration Bus to other address controllers on other boards. Each leaf arbitrator has four prioritized request in lines, each such line being coupled to a single address controller serviced by that leaf arbitrator. By default, each leaf arbitrator and the top arbitrator implement a prioritized algorithm.

Abstract: Snooping is implemented on a split transaction snooping bus for a computer system having one or many such buses. Circuit boards including CPU or other devices and/or distributed memory, data input/output buffers, queues including request tag queues, coherent input queues ("CIQ"), and address controller implementing address bus arbitration plug-into one or more split transaction snooping bus systems. All devices snoop on the address bus to learn whether an identified line is owned or shared, and an appropriate owned/shared signal is issued. Receipt of an ignore signal blocks CIQ loading of a transaction until the transaction is reloaded and ignore is deasserted. Ownership of a requested memory line transfers immediately at time of request. Asserted requests are queued such that state transactions on the address bus occur atomically logically without dependence upon the request. Subsequent requests for the same data are tagged to become the responsibility of the owner-requestor.

Abstract: A split transaction snooping bus protocol and architecture is provided for use in a system having one or many such buses. Circuit boards including CPU or other devices and/or distributed memory, data input/output buffers, queues including request tag queues, coherent input queues ("CIQ"), and address controller implementing address bus arbitration plug-into one or more split transaction snooping bus systems. All devices snoop on the address bus to learn whether an identified line is owned or shared, and an appropriate owned/shared signal is issued. Receipt of an ignore signal blocks CIQ loading of a transaction until the transaction is reloaded and ignore is deasserted. Ownership of a requested memory line transfers immediately at time of request. Asserted requests are queued such that state transactions on the address bus occur atomically logically without dependence upon the request. Subsequent requests for the same data are tagged to become the responsibility of the owner-requestor.

Abstract: In a multiprocessor system, a method, apparatus, and article of manufacture for maintaining the proper sequence of store/write operations between multiple processors to remote I/O devices without requiring changes to application software. A synchronizer is employed to synchronize write operations to the remote I/O device, and the write operations are synchronized individually upon detection and emulation, or as a group upon detection of the release of a mutual exclusion lock.

Abstract: In a computer system implementing state transitions that change logically and atomically at an address packet independently of a response, the coherence domain is extended among distributed memory. As such, memory line ownership transfers upon request, and not upon requestor receipt of data. Requestor receipt of data is rapidly implemented by providing a ReadToShareFork transaction that simultaneously causes a write-type operation that updates invalid data from a requested memory address, and provides the updated data to the requesting device. More specifically, when writing valid data to memory, the ReadToShare Fork transaction simultaneously causes reissuance of the originally requested transaction using the same memory address and ID information. The requesting device upon recognizing its transaction ID on the bus system will pull the now valid data from the desired memory location.

Abstract: An architecture for a multiprocessor computer system is provided. The multiprocessor computer system includes multiple repeater nodes. Each repeater node includes a transaction repeater and at least one bus device coupled to the repeater on a lower level bus. The repeater nodes are connected by an upper level bus. Each bus device includes an incoming queue. Transaction originating in a particular repeater node are stored in the incoming queue, whereas transactions originating in other repeater nodes bypass the incoming queue to the bus device. A control signal is asserted by the repeater so that a transaction is received by bus devices in the originating node from the incoming queues at the same time and in the same order it is received by bus devices in non-originating nodes. Thus a hierarchical bus structure is provided that overcomes physical/electrical limitations of single bus architecture while maximizing bus bandwidth utilization.

Abstract: The present invention provides a cache manager (CM) for use with an address translation table (ATT) which take advantage of way information, available when a cache line is first cached, for efficiently accessing a multi-way cache of a computer system having a main memory and one or more processors. The main memory and the ATT are page-oriented while the cache is organized using cache lines. The cache includes a plurality of cache lines divided into a number of segments corresponding to the number of "ways". Each cache line includes an address tag (AT) field and a data field. The way information is stored in the ATT for later cache access. In this implementation, "waylets" provide an efficiency mechanism for storing the way information whenever a cache line is cached. Accordingly, each table entry of the ATT includes a virtual address (VA) field, a physical address (PA) field, and a plurality of waylets associated with each pair of VA and PA fields.

Abstract: In a multiprocessor system having a shared memory, each central processor services copyback requests from other central processors. Each central processor has a writeback buffer along with a plurality of tag buffers and an associated snoop architecture for processing writeback and copyback commands. Each central processor includes a cache subsystem having a system interface, a main cache and an associated tag array. The system interface has an address controller and data controller, each having separate input and output queues for interfacing between the central processor and system control and data buses. The address controller includes a set of duplicate tags that mirror the tags associated with the main cache, and an auxiliary tag input buffer and auxiliary tag output buffer. The address controller has for each line in the output queue an associated pointer that indicates the location in the data controller where data is stored that is associated with output queued commands.

Abstract: The present invention provides an "alert" interface for a component which can be safely "hot-plugged/unplugged" to an "alert" interconnect of an electrically powered system. The alert interface has a mating edge which includes daughter precharge/ground connectors, a daughter (engage) waning connector, a number of daughter signal connectors and a daughter engage connector. The alert interconnect includes corresponding mother connectors. The respective connectors of the interconnect and the interface are arranged so that they mate in the following exemplary order when the interface is hot-plugged/unplugged to the interconnect: precharge/ground connectors, warning connectors, signal connectors and finally engage connectors. When the daughter (engage) warning connector mates with the mother warning connector, the component sends an "engage warning" signal to the powered system.