Abstract: A multiprocessor data processing system requires careful management to maintain cache coherency. In conventional systems using a MESI approach, two or more processors will often compete for ownership of a common cache line. As a result, ownership of the cache line will frequently “bounce” between multiple processors, which causes a significant reduction in cache efficiency. The preferred embodiment provides a modified MESI state which holds the status of the cache line static for a fixed period of time, which eliminates the bounce effect from contention between multiple processors.

Abstract: A multiprocessor data processing system requires careful management to maintain cache coherency. Conventional systems using a MESI approach sacrifice some performance with inefficient lock-acquisition and lock-retention techniques. The disclosed system provides additional cache states, indicator bits, and lock-acquisition routines to improve cache performance. The additional cache states allow cache state transition sequences to be optimized by replacing frequently-occurring and inefficient MESI code sequences with improved sequences using modified cache states.

Abstract: A multiprocessor data processing system requires careful management to maintain cache coherency. Conventional systems using a MESI approach sacrifice some performance with inefficient lock-acquisition and lock-retention techniques. The disclosed system provides additional cache states, indicator bits, and lock-acquisition routines to improve cache performance. The additional cache states allow cache state transition sequences to be optimized. In particular, the claimed system and method provides that a given processor, after acquiring a lock or reservation to a given cache line, will keep the lock, to make successive modifications to the cache line, instead of releasing it to other processors after making only one modification. By doing so, the overhead typically required to acquire a lock before making any cache line modification is eliminated for successive modifications.

Abstract: A multiprocessor data processing system requires careful management to maintain cache coherency. Conventional systems using a MESI approach sacrifice some performance with inefficient lock-acquisition and lock-retention techniques. The disclosed system provides additional cache states, indicator bits, and lock-acquisition routines to improve cache performance. In particular, as multiple processors compete for the same cache line, a significant amount of processor time is lost determining if another processor's cache line lock has been released and attempting to reserve that cache line while it is still owned by the other processor. The preferred embodiment provides an indicator bit with the cache store command which specifically indicates whether the store also acts as a lock-release.

Abstract: Disclosed is a method of operation within a processor that permits load instructions to be issued speculatively. An instruction sequence is received that includes multiple barrier instructions and a load instruction that follows the barrier instructions in the instruction sequence. In response to the multiple barrier instructions, barrier operations are issued on an interconnect coupled to the processor. Also, while the barrier operations are pending, a load request associated with the load instruction is speculatively issued. When the load request is issued, a flag is set to indicate that it was speculatively issued. The flag is reset when acknowledgments of all the barrier operations are received. Data that is returned before the acknowledgments are received is temporarily held and forwarded to the register and/or execution unit of the processor only after the acknowledgments are received.

Abstract: A non-uniform memory access (NUMA) computer system includes a remote node coupled by a node interconnect to a home node having a home system memory. The remote node includes a local interconnect, a processing unit and at least one cache coupled to the local interconnect, and a node controller coupled between the local interconnect and the node interconnect. The processing unit first issues, on the local interconnect, a read-type request targeting data resident in the home system memory with a flag in the read-type request set to a first state to indicate only local servicing of the read-type request. In response to inability to service the read-type request locally in the remote node, the processing unit reissues the read-type request with the flag set to a second state to instruct the node controller to transmit the read-type request to the home node.

Abstract: A method of handling a write operation in a multiprocessor computer system wherein each processing unit has a respective cache, by determining that a new value for a store instruction is the same as a current value already contained in the memory hierarchy, and discarding the store instruction without issuing any associated cache operation in response to this determination. When a store hit occurs, the current value is retrieved from the local cache. When a store miss occurs, the current value is retrieved from a remote cache by issuing a read request. The comparison may be performed using a portion of the cache line which is less than a granule size of the cache line. A store gathering queue can be use to collect pending store instructions that are directed to different portions of the same cache line.

Abstract: A method of handling a write operation in a multiprocessor computer system wherein each processing unit has a respective cache, by determining that a new value for a store instruction is the same as a current value already contained in the memory hierarchy, and discarding the store instruction without issuing any associated cache operation in response to this determination. When a store hit occurs, the current value is retrieved from the local cache. When a store miss occurs, the current value is retrieved from a remote cache by issuing a read request. The comparison may be performed using a portion of the cache line which is less than a granule size of the cache line. A store gathering queue can be use to collect pending store instructions that are directed to different portions of the same cache line.

Abstract: Disclosed is a multiprocessor data processing system that executes loads transactions out of order with respect to a barrier operation. The data processing system includes a memory and a plurality of processors coupled to an interconnect. At least one of the processors includes an instruction sequencing unit for fetching an instruction sequence in program order for execution. The instruction sequence includes a first and a second load instruction and a barrier instruction, which is between the first and second load instructions in the instruction sequence. Also included in the processor is a load/store unit (LSU), which has a load request queue (LRQ) that temporarily buffers load requests associated with the first and second load instructions. The LRQ is coupled to a load request arbitration unit, which selects an order of issuing the load requests from the LRQ.

Abstract: Disclosed is a method of operating a processor, by which a speculatively issued load request, which fetches incorrect data, is recycled. An instruction sequence, which includes a barrier instruction and a load instruction that follows the barrier instruction in program order, is received for execution. In response to the barrier instruction, a barrier operation is issued on an interconnect. Following, in response to the load instruction and while the barrier operation is pending, a load request is issued to memory. When a pre-determined type of invalidate, which is affiliated with the load request, is received before the receipt of an acknowledgment for the barrier operation, data that is returned by memory in response to the load request is discarded and the load request is re-issued. The pre-determined type of invalidate includes, for example, a snoop invalidate.

Abstract: In addition to an address tag, a coherency state and an LRU position, each cache directory entry includes historical processor access and snoop operation information for the corresponding cache line. The historical processor access and snoop operation information includes different subentries for each different processor which has accessed the corresponding cache line, with subentries being “pushed” along the stack when a new processor accesses the subject cache line. Each subentries contains the processor identifier for the corresponding-processor which accessed the cache line, a processor access history segment, and a snoop operation history segment. The processor access history segment contains one or more opcodes identifying the operations which were performed by the processor, and timestamps associated with each opcode.

Abstract: A method and system for scarfing data during a data access transaction within a hierarchical data storage system. A data access request is delivered from a source device to a plurality of data storage devices. The access request includes a target address and a source path tag, wherein the source path tag includes a device identification tag that uniquely identifies a data storage device within a given level of the system traversed by the access request. A device identification tag that uniquely identifies the third party transactor within a given memory level is appended to the source path tag such that the third party transactor can scarf returning data without reserving a scarf queue entry.

Abstract: A method and system for managing a data access transaction within a hierarchical data storage system. In accordance with the method of the present invention, a data access request is delivered from a source device to multiple data storage devices within the hierarchical data storage system. The data access request includes a source path tag and a target address. At least one device identification tag is added to the source path tag, wherein the at least one device identification tag uniquely identifies a data storage device within each level of the hierarchical data storage system traversed by the data access request such that the data access transaction can be processed in accordance with source path information that is incrementally encoded within the data access request as the data access request traverses the hierarchical data storage system.

Abstract: A method and system for processing a split data access transaction within a hierarchical data storage system. In accordance with the method of the present invention, a data access request is delivered from a source device onto an address bus that is shared by a plurality of data storage devices within the hierarchical data storage system, wherein the data access request includes a target address and a source path tag. The source path tag includes at least one device identification tag that uniquely identifies at least one data storage device within each level of the hierarchical data storage system traversed by the data access request. In response to a data access request hit at a given data storage device, a data access response is delivered onto a data bus, wherein the data access response includes the source path tag and the target address.

Abstract: A set-associative cache memory having asymmetric latency among sets is disclosed. The cache memory has multiple congruence classes of cache lines. Each congruence class includes a number of sets organized in a set-associative manner. The cache memory further includes a means for accessing at least one of the sets faster than the remaining sets having an identical access latency.

Abstract: A method of maintaining coherency in a cache hierarchy of a processing unit of a computer system, wherein the upper level (L1) cache includes a split instruction/data cache. In one implementation, the L1 data cache is store-through, and each processing unit has a lower level (L2) cache. When the lower level cache receives a cache operation requiring invalidation of a program instruction in the L1 instruction cache (i.e., a store operation or a snooped kill), the L2 cache sends an invalidation transaction (e.g., icbi) to the instruction cache. The L2 cache is fully inclusive of both instructions and data. In another implementation, the L1 data cache is write-back, and a store address queue in the processor core is used to continually propagate pipelined address sequences to the lower levels of the memory hierarchy, i.e., to an L2 cache or, if there is no L2 cache, then to the system bus. If there is no L2 cache, then the cache operations may be snooped directly against the L1 instruction cache.

Abstract: A method of maintaining coherency in a multiprocessor computer system wherein each processing unit's cache has sectored cache lines. A first cache coherency state is assigned to one of the sectors of a particular cache line, and a second cache coherency state, different from the first cache coherency state, is assigned to the overall cache line while maintaining the first cache coherency state for the first sector. The first cache coherency state may provide an indication that the first sector contains a valid value which is not shared with any other cache (i.e., an exclusive or modified state), and the second cache coherency state may provide an indication that at least one of the sectors in the cache line contains a valid value which is shared with at least one other cache (a shared, recently-read, or tagged state). Other coherency states may be applied to other sectors in the same cache line.

Abstract: In response to a need to initiate a global operation, a bus master within a multiprocessor system issues a combined token and operation request on a bus coupled to the bus master. The combined token and operation request solicits a token required to complete the global operation and identifies the global operation to be processed with the token, if granted. Upon receiving a combined response acknowledging both the token and operation portions of the combined request, the bus master treats the global operation as complete. If a combined response acknowledging the token portion of the combined request but retrying the operation portion (i.e., at least one snooper is busy processing a previous global operation), the bus master issues an operation request (only) for the operation portion of the combined request.

Abstract: A method of maintaining coherency in a multiprocessor computer system wherein each processing unit's cache has sectored cache lines. A first cache coherency state is assigned to one of the sectors of a particular cache line, and a second cache coherency state, different from the first cache coherency state, is assigned to the overall cache line while maintaining the first cache coherency state for the first sector. The first cache coherency state may provide an indication that the first sector contains a valid value which is not shared with any other cache (i.e., an exclusive or modified state), and the second cache coherency state may provide an indication that at least one of the sectors in the cache line contains a valid value which is shared with at least one other cache (a shared, recently-read, or tagged state). Other coherency states may be applied to other sectors in the same cache line.

Abstract: A multiprocessor data processing system requires careful management to maintain cache coherency. Conventional systems using a MESI approach sacrifice some performance with inefficient lock-acquisition and lock-retention techniques. The disclosed system provides additional cache states, indicator bits, and lock-acquisition routines to improve cache performance. In particular, as multiple processors compete for the same cache line, a significant amount of processor time is lost determining if another processor's cache line lock has been released and attempting to reserve that cache line while it is still owned by the other processor. The preferred embodiment provides an additional cache state which specifically indicates that a processor has released its lock on a cache line after it has performed any necessary modifications.