Intel Corporation

International Business Machines Corporation

Wen-Hann Wang

Linda J Rankin

Kai Cheng

Nitin Y Borkar

Paul R Pierce

Gregory E Dermer

Richard H Hofsheier

Nitin V Sarangdhar

Konrad K Lai

Gurbir Singh

John M Bauer

Michael W Rhodehamel

Scott H Robinson

Matthew A Fisch

Mandar S Joshi

Jih-Kwon Peir

Steve Y Zhang

Ashwani Kumar Gupta

Kimming K So

Quinn W Merrell

A multiprocessor system comprises at least one processing module, at least one I/O module, and an interconnect network to connect the at least one processing module with the at least one input/output module. In an example embodiment, the interconnect network comprises at least two bridges to send and receive transactions between the input/output modules and the processing module. The interconnect network further comprises at least two crossbar switches to route the transactions over a high bandwidth switch connection. Using embodiments of the interconnect network allows high bandwidth communication between processing modules and I/O modules. Standard processing module hardware can be used with the interconnect network without modifying the BIOS or the operating system. Furthermore, using the interconnect network of embodiments of the present invention is non-invasive to the processor motherboard. The processor memory bus, clock, and reset logic all remain intact.

Scalable distributed memory and I/O multiprocessor system

A multiprocessor system comprises at least one processing module, at least one I/O module, and an interconnect network to connect the at least one processing module with the at least one input/output module. In an example embodiment, the interconnect network comprises at least two bridges to send and receive transactions between the input/output modules and the processing module. The interconnect network further comprises at least two crossbar switches to route the transactions over a high bandwidth switch connection. Using embodiments of the interconnect network allows high bandwidth communication between processing modules and I/O modules. Standard processing module hardware can be used with the interconnect network without modifying the BICS or the operating system. Furthermore, using the interconnect network of embodiments of the present invention is non-invasive to the processor motherboard. The processor memory bus, clock, and reset logic all remain intact.

Scalable memory and I/O multiprocessor systems

Scalable distributed memory and I/O multiprocessor systems and associated methods

A multiprocessor system comprises at least one processing module, at least one I/O module, and an interconnect network to connect the at least one processing module with the at least one input/output module. In an example embodiment, the interconnect network comprises at least two bridges to send and receive transactions between the input/output modules and the processing module The interconnect network further comprises at least two crossbar switches to route the transactions over a high bandwidth switch connection. Using embodiments of the interconnect network allows high bandwidth communication between processing modules and I/O modules Standard processing module hardware can be used with the interconnect network without modifying the BIOS or the operating system. Furthermore, using the interconnect network of embodiments of the present invention is non-invasive to the processor motherboard. The processor memory bus, clock, and reset logic all remain intact.

The invention provides a cache management system comprising in various embodiments pre-load and pre-own functionality to enhance cache efficiency in shared memory distributed cache multiprocessor computer systems. Some embodiments of the invention comprise an invalidation history table to record the line addresses of cache lines invalidated through dirty or clean invalidation, and which is used such that invalidated cache lines recorded in an invalidation history table are reloaded into cache by monitoring the bus for cache line addresses of cache lines recorded in the invalidation history table. In some further embodiments, a write-back bit associated with each L2 cache entry records when either a hit to the same line in another processor is detected or when the same line is invalidated in another processor's cache, and the system broadcasts write-backs from the selected local cache only when the line being written back has a write-back bit that has been set.

Cache line pre-load and pre-own based on cache coherence speculation

In a computer system, an apparatus for handling lock conditions wherein a first instruction executed by a first processor processes data that is common to a second processor while the second processor is locked from simultaneously executing a second instruction that also processes this same data. A lock bit is set when the first processor begins execution of the first instruction. Thereupon, the second processor is prevented from executing its instruction until the first processor has completed its processing of the shared data. Hence, the second processor queues its request in a buffer. The lock bit is cleared after the first processor has completed execution of its instruction. The first processor then checks the buffer for any outstanding requests. In response to the second processor's queued request, the first processor transmits a signal to the second processor indicating that the data is now not locked.

Apparatus and method for caching lock conditions in a multi-processor

The present invention includes a computer system having an on-processor predictor tag array, an off-processor cache memory, and comparison circuitry. The on-processor predictor tag array contains first portions of tag information for multiple ways and multiple sets. The off-processor cache memory includes memory locations to store data and second portions of tag information. The comparison circuitry makes a first comparison of a first portion of an address with the first portions of tag information for the ways of one of the sets and uses results of the first comparison in predicting which of the ways, if any, correspond to the address. The comparison circuitry also makes a second comparison of the second portion of the address with sections of the second portions of tag information identified by the predicted way and the address.

Computer system having tag information in a processor and cache memory

A system and method for reducing the number of writeback operations performed by level two (L2) or higher level cache memories in a microprocessor system having an integrated hierarchical cache structure. Writeback operations of modified victim lines in L2 or higher level caches are cancelled if an associated cache line, having a 'modified' status, is located in a lower level cache. In one embodiment of the present invention, writeback operations of modified victim lines in L2 or higher level caches are also cancelled if an associated cache line, having a 'clean' status, is located in a lower level cache.

Backward inquiry to lower level caches prior to the eviction of a

A method and apparatus for cache memory replacement line identification have a cache interface which provides a communication interface between a cache memory and a controller for the cache memory. The interface includes an address bus, a data bus, and a status bus. The address bus transfers requested addresses from the controller to the cache memory. The data bus transfers data associated with requested addresses from the controller to the cache memory, and also transfers replacement line addresses from the cache memory to the controller. The status bus transfers status information associated with the requested addresses from the cache memory to the controller which indicate whether the requested addresses are contained in the cache memory. In one embodiment, the data bus also transfers cache line data associated with a requested address from the cache memory to the controller when the requested address hits the cache memory.

Method and apparatus for cache memory replacement line identification

A computer system comprising a plurality of caching agents with a cache hierarchy, the caching agents sharing memory across a system bus and issuing memory access requests in accordance with a protocol wherein a line of a cache has a present state comprising one of a plurality of line states. The plurality of line states includes a modified (M) state, wherein a line of a first caching agent in M state has data which is more recent than any other copy in the system; an exclusive (E) state, wherein a line in E state in a first caching agent is the only one of the agents in the system which has a copy of the data in a line of the cache, the first caching agent modifying the data in the cache line independent of other said agents coupled to the system bus; a shared (S) state, wherein a line in S state indicates that more than one of the agents has a copy of the data in the line; and an invalid (I) state indicating that the line does not exist in the cache. A read or a write to a line in I state results in a cache miss. The present invention associates states with lines and defines rules governing state transitions. State transitions depend on both processor generated activities and activities by other bus agents, including other processors. Data consistency is guaranteed in systems having multiple levels of cache and shared memory and/or multiple active agents, such that no agent ever reads stale data and actions are serialized as needed.

Apparatus for maintaining multilevel cache hierarchy coherency in a

A method and apparatus for transferring information between a processor and a memory system utilizing a chunk write buffer, where read and write requests to the L2 cache memory are controlled by the processor. The cache line associated with each such request is larger than the interface coupling the L2 cache memory and the processor. Read requests are returned from the L2 cache memory to the processor in burst fashion. Write requests are transferred from the processor to the L2 cache memory during clock cycles in which the processor does not require the interface for a read request. Write requests need not be transferred in burst fashion; rather, a portion of the write request corresponding to the size of the interface, referred to as a chunk, is transferred from the processor to the L2 cache memory and stored temporarily in the chunk write buffer. When the processor has transferred the entire cache line to the L2 cache memory, the processor signals the L2 cache memory to transfer the contents of the chunk write buffer into the data array of the cache memory.

Method and apparatus for transferring information between a processor

A cache memory with reduced request-blocking blocks requests from being accepted by the cache memory based on the types of requests the cache is already servicing. A request which hits the cache memory or a request which misses the cache memory but does not conflict with any requests already being serviced is not blocked. A request which misses the cache memory and also conflicts with a request(s) already being serviced causes the request to be blocked. In one embodiment, conflicts for write requests are determined by checking whether the cache is already retrieving a cache line from system memory for a request which maps into the same cache set as the write request. If such a request exists, then a conflict occurs. In this embodiment, conflicts for read requests are determined by checking whether the cache is already servicing an outstanding request to memory for the same address. If so, then a conflict occurs. If not, then a conflict does not occur unless the victim line for the read request is dirty and no space exists in a write-back buffer to temporarily store the victim line.

Growing community of inventors

Portland, OR, United States of America

Wen-Hann Wang