Abstract: A set associative cache includes a cache controller, a directory, and an array including at least one congruence class containing a plurality of sets. The plurality of sets are partitioned into multiple groups according to which of a plurality of information types each set can store. The sets are partitioned so that at least two of the groups include the same set and at least one of the sets can store fewer than all of the information types. The cache controller then implements different cache policies for at least two of the plurality of groups, thus permitting the operation of the cache to be individually optimized for different information types.

Abstract: A method of operating a processing unit of a computer system, by issuing an instruction having an explicit prefetch request directly from an instruction sequence unit to a prefetch unit of the processing unit. The invention applies to values that are either operand data or instructions. In a preferred embodiment, two prefetch units are used, the first prefetch unit being hardware independent and dynamically monitoring one or more active streams associated with operations carried out by a core of the processing unit, and the second prefetch unit being aware of the lower level storage subsystem and sending with the prefetch request an indication that a prefetch value is to be loaded into a lower level cache of the processing unit. The invention may advantageously associate each prefetch request with a stream ID of an associated processor stream, or a processor ID of the requesting processing unit (the latter feature is particularly useful for caches which are shared by a processing unit cluster).

Abstract: A method of operating a processing unit of a computer system, by issuing an instruction having an explicit prefetch request directly from an instruction sequence unit to a prefetch unit of the processing unit. The invention applies to values that are either operand data or instructions. In a preferred embodiment, two prefetch units are used, the first prefetch unit being hardware independent and dynamically monitoring one or more active streams associated with operations carried out by a core of the processing unit, and the second prefetch unit being aware of the lower level storage subsystem and sending with the prefetch request an indication that a prefetch value is to be loaded into a lower level cache of the processing unit. The invention may advantageously associate each prefetch request with a stream ID of an associated processor stream, or a processor ID of the requesting processing unit (the latter feature is particularly useful for caches which are shared by a processing unit cluster).

Abstract: A data processing system includes a processing unit, a distributed memory including a local memory and a remote memory having differing access latencies, and a cache coupled to the processing unit and to the distributed memory. The cache includes a congruence class containing a plurality of cache lines and a plurality of latency indicators that each indicate an access latency to the distributed memory for a respective one of the cache lines. The cache further includes a cache controller that selects a cache line in the congruence class as a castout victim in response to the access latencies indicated by the plurality of latency indicators. In one preferred embodiment, the cache controller preferentially selects as castout victims cache lines having relatively short access latencies.

Abstract: A data processing system includes an interconnect, a system memory and a number of snoopers coupled to the interconnect, and response logic. In response to a requesting snooper issuing a data request on the interconnect specifying a memory address, the snoopers provide snoop responses. The response logic compiles the snoop responses to obtain a combined response including an indication of a demand-source snooper that will source requested data associated with the memory address to the requesting snooper and an indication of whether additional non-requested data will be supplied to the requesting snooper. This combined response is then transmitted to the snoopers on the interconnect to direct the provision of the requested data, and possibly unrequested prefetch data, to the requesting snooper.

Abstract: A multiprocessor data processing system includes an interconnect, a plurality of processing units coupled to the interconnect, and at least one system memory and a plurality of caches coupled to the plurality of processing units. A cache suitable for use in such a data processing system includes data storage containing multiple granules of data and a number of state fields associated with the granules of data. Each state field has a plurality of possible states including an O state indicating that an associated granule is consistent with corresponding data in the memory and has unknown coherency with respect to peer caches in the data processing system. Thus, a cache is permitted to store memory-consistent, but possibly non-coherent data in order to offer processing units in the data processing system lower latency to an image of system memory.

Abstract: A method of operating a computer system is disclosed in which an instruction having an explicit prefetch request is issued directly from an instruction sequence unit to a prefetch unit of a processing unit. In a preferred embodiment, two prefetch units are used, the first prefetch unit being hardware independent and dynamically monitoring one or more active streams associated with operations carried out by a core of the processing unit, and the second prefetch unit being aware of the lower level storage subsystem and sending with the prefetch request an indication that a prefetch value is to be loaded into a lower level cache of the processing unit. The invention may advantageously associate each prefetch request with a stream ID of an associated processor stream, or a processor ID of the requesting processing unit (the latter feature is particularly useful for caches which are shared by a processing unit cluster).

Abstract: A cache coherency protocol uses a “Exclusive-Deallocate” (ED) coherency state to indicate that a particular value is currently held in an upper level cache in an exclusive, unmodified form (not shared with any other caches of the computer system, including caches associated with the same processing unit), so that the value can conveniently be modified without any lower level bus transactions since no lower level caches have allocated a line for the value. If the value is subsequently modified in the upper level cache, its coherency state is simply switched to “modified” without the need for any bus transactions. Conversely, if the value is evicted from the upper level cache without ever having been modified, it can be loaded into the lower level cache with a coherency state indicating that the lower level cache contains the unmodified value exclusive of all other caches in other processing units of the computer system.

Abstract: A method and system for maintaining allocation information on data castout from an upper level cache provides a cache control with the ability to select victims based on whether a cache entry is present due to a read request from a higher level in the memory hierarchy or is present due to being modified in the higher level and then castout to the lower level. The information maintained may be a single bit indicating this status, may be a separate least-recently-used (LRU) array value indicating the order of allocation in the lower level for storage of cache entries castout from the higher level.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (MU) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the MU state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The MU state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the MU state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: A method and system for allocating lower level cache entries for data castout from an upper level cache provides improved computer system performance by adjusting the ordering of least-recently-used (LRU) information within a cache. Data that is castout from a higher level cache can be written after a read is satisfied and the castout entry will not be labeled as most-recently-used. This improves performance under certain operating conditions of a computing system, as castout data is often less important to keep in lower level cache than data that is also present in the higher level cache.

Abstract: A method of operating a computer system is disclosed in which an instruction having an explicit prefetch request is issued directly from an instruction sequence unit to a prefetch unit of a processing unit. In a preferred embodiment, two prefetch units are used, the first prefetch unit being hardware independent and dynamically monitoring one or more active streams associated with operations carried out by a core of the processing unit, and the second prefetch unit being aware of the lower level storage subsystem and sending with the prefetch request an indication that a prefetch value is to be loaded into a lower level cache of the processing unit. The invention may advantageously associate each prefetch request with a stream ID of an associated processor stream, or a processor ID of the requesting processing unit (the latter feature is particularly useful for caches which are shared by a processing unit cluster).

Abstract: A multiprocessor data processing system includes an interconnect, a plurality of processing units coupled to the interconnect, and at least one system memory and a plurality of caches coupled to the plurality of processing units. A cache suitable for use in such a data processing system includes data storage containing a data granule, a state field associated with the data granule, and a cache controller. The state field has a plurality of possible states including a first state that indicates that the data granule is consistent with corresponding data in the memory and has unknown coherency with respect to other peer caches among the plurality of caches. To update the state of the data granule from the first state, the cache controller issues on the interconnect a transaction specifying an address associated with the data granule. In response to receipt of a combined response of the plurality of caches, the cache controller updates the state field to a second state among the plurality of possible states.

Abstract: A method and system for bypassing cache levels when storing data castout from an upper level cache provides a memory hierarchy that can selectively skip one more more intermediate levels when writing castout entries from a higher level cache based on a number of detected conditions. The intermediate levels may be bypassed when an intermediate cache level is busy, has an entry with an address conflict with the castout value, or may skip levels based on program control. The control providing the skipping selection may be driven by a detector that analyzes load/store operations of a processor in order to produce efficient operation under changing memory use conditions.

Abstract: A multiprocessor data processing system includes an interconnect, a plurality of processing units coupled to the interconnect, and at least one system memory and a plurality of caches coupled to the plurality of processing units. A cache suitable for use in such a data processing system includes data storage containing multiple data granules and a number of state fields associated with the data granules. Each state field has a plurality of possible states including an OR state that indicates that an associated granule is consistent with corresponding data in the memory, that the associated data granule has unknown coherency with respect to other peer caches in the data processing system, and that the cache is responsible, among all of its peer caches that may store the associated data granule in a memory-consistent state with unknown coherency, for sourcing the data granule in response to a request.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (MU) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the MU state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The MU state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the MU state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: A multiprocessor data processing system includes an interconnect, a plurality of processing units coupled to the interconnect, and at least one system memory and a plurality of caches coupled to the plurality of processing units. A cache suitable for use in such a data processing system includes data storage containing multiple granules of data and a number of state fields associated with the granules of data. Each state field has a plurality of possible states including an O state indicating that an associated granule is consistent with corresponding data in the memory and has unknown coherency with respect to peer caches in the data processing system. The cache updates the state field from the O state to another of the plurality of states in response to a snooped transaction on the interconnect.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (Mu) cache state to indicate that a value held in 5 a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the Mu state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The Mu state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the Mu state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (MU) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the MU state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The MU state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the MU state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.

Abstract: A novel cache coherency protocol provides a modified-unsolicited (Mu) cache state to indicate that a value held in a cache line has been modified (i.e., is not currently consistent with system memory), but was modified by another processing unit, not by the processing unit associated with the cache that currently contains the value in the Mu state, and that the value is held exclusive of any other horizontally adjacent caches. Because the value is exclusively held, it may be modified in that cache without the necessity of issuing a bus transaction to other horizontal caches in the memory hierarchy. The Mu state may be applied as a result of a snoop response to a read request. The read request can include a flag to indicate that the requesting cache is capable of utilizing the Mu state. Alternatively, a flag may be provided with intervention data to indicate that the requesting cache should utilize the modified-unsolicited state.