Abstract: A system and method are provided for performing setup operations for receiving a different amount of data while processors are performing message passing interface (MPI) tasks. Mechanisms for adjusting the balance of processing workloads of the processors are provided so as to minimize wait periods for waiting for all of the processors to call a synchronization operation. An MPI load balancing controller maintains a history that provides a profile of the tasks with regard to their calls to synchronization operations. From this information, it can be determined which processors should have their processing loads lightened and which processors are able to handle additional processing loads without significantly negatively affecting the overall operation of the parallel execution system. As a result, setup operations may be performed while processors are performing MPI tasks to prepare for receiving different sized portions of data in a subsequent computation cycle based on the history.

Abstract: A mechanism for assigning memory to on-chip cache coherence domains assigns caches within a processing unit to coherence domains. The mechanism assigns chunks of memory to the coherence domains. The mechanism monitors applications running on cores within the processing unit to identify needs of the applications. The mechanism may then reassign memory chunks to the cache coherence domains based on the needs of the applications running in the coherence domains. When a memory controller receives the cache miss, the memory controller may look up the address in a lookup table that maps memory chunks to cache coherence domains. Snoop requests are sent to caches within the coherence domain. If a cache line is found in a cache within the coherence domain, the cache line is returned to the originating cache by the cache containing the cache line either directly or through the memory controller.

Abstract: A mechanism is provided in a cache for providing a read and write aware cache. The mechanism partitions a large cache into a read-often region and a write-often region. The mechanism considers read/write frequency in a non-uniform cache architecture replacement policy. A frequently written cache line is placed in one of the farther banks. A frequently read cache line is placed in one of the closer banks. The size ratio between read-often and write-often regions may be static or dynamic. The boundary between the read-often region and the write-often region may be distinct or fuzzy.

Abstract: In at least one embodiment, a processor detects during execution of program code whether a load instruction within the program code is associated with a hint. In response to detecting that the load instruction is not associated with a hint, the processor retrieves a full cache line of data from the memory hierarchy into the processor in response to the load instruction. In response to detecting that the load instruction is associated with a hint, a processor retrieves a partial cache line of data into the processor from the memory hierarchy in response to the load instruction.

Abstract: Mechanisms are provided for inhibiting precharging of memory cells of a dynamic random access memory (DRAM) structure. The mechanisms receive a command for accessing memory cells of the DRAM structure. The mechanisms further determine, based on the command, if precharging the memory cells following accessing the memory cells is to be inhibited. Moreover, the mechanisms send, in response to the determination indicating that precharging the memory cells is to be inhibited, a command to blocking logic of the DRAM structure to block precharging of the memory cells following accessing the memory cells.

Abstract: According to a method of data processing, a memory controller receives a prefetch load request from a processor core of a data processing system. The prefetch load request specifies a requested line of data. In response to receipt of the prefetch load request, the memory controller determines by reference to a stream of demand requests how much data is to be supplied to the processor core in response to the prefetch load request. In response to the memory controller determining to provide less than all of the requested line of data, the memory controller provides less than all of the requested line of data to the processor core.

Abstract: Mechanisms are provided for performing setup operations for receiving a different amount of data while processors are performing message passing interface (MPI) tasks. Mechanisms for adjusting the balance of processing workloads of the processors are provided so us to minimize wait periods for waiting for all of the processors to call a synchronization operation. An MPI load balancing controller maintains a history that provides a profile of the tasks with regard to their calls to synchronization operations. From this information, it can be determined which processors should have their processing loads lightened and which processors are able to handle additional processing loads without significantly negatively affecting the overall operation of the parallel execution system. As a result, setup operations may be performed while processors are performing MPI tasks to prepare for receiving different sized portions of data in a subsequent computation cycle based on the history.

Abstract: A technique for data prefetching using indirect addressing includes monitoring data pointer values, associated with an array, in an access stream to a memory. The technique determines whether a pattern exists in the data pointer values. A prefetch table is then populated with respective entries that correspond to respective array address/data pointer pairs based on a predicted pattern in the data pointer values. Respective data blocks (e.g., respective cache lines) are then prefetched (e.g., from the memory or another memory) based on the respective entries in the prefetch table.

Abstract: A method is provided for implementing a multi-tiered full-graph interconnect architecture. In order to implement a multi-tiered full-graph interconnect architecture, a plurality of processors are coupled to one another to create a plurality of processor books. The plurality of processor books are coupled together to create a plurality of supernodes. Then, the plurality of supernodes are coupled together to create the multi-tiered full-graph interconnect architecture. Data is then transmitted from one processor to another within the multi-tiered full-graph interconnect architecture based on an addressing scheme that specifies at least a supernode and a processor book associated with a target processor to which the data is to be transmitted.

Abstract: A scalable space-optimized and energy-efficient computing system is provided. The computing system comprises a plurality of modular compartments in at least one level of a frame configured in a hexadron configuration. The computing system also comprises an air inlet, an air mixing plenum, and at least one fan. In the computing system the plurality of modular compartments are affixed above the air inlet, the air mixing plenum is affixed above the plurality of modular compartments, and the at least one fan is affixed above the air mixing plenum. When at least one module is inserted into one of the plurality of modular compartments, the module couples to a backplane within the frame.

Abstract: A technique for performing indirect data prefetching includes determining a first memory address of a pointer associated with a data prefetch instruction. Content of a memory at the first memory address is then fetched. A second memory address is determined from the content of the memory at the first memory address. Finally, a data block (e.g., a cache line) including data at the second memory address is fetched (e.g., from the memory or another memory).

Abstract: A processor includes a first address translation engine, a second address translation engine, and a prefetch engine. The first address translation engine is configured to determine a first memory address of a pointer associated with a data prefetch instruction. The prefetch engine is coupled to the first translation engine and is configured to fetch content, included in a first data block (e.g., a first cache line) of a memory, at the first memory address. The second address translation engine is coupled to the prefetch engine and is configured to determine a second memory address based on the content of the memory at the first memory address. The prefetch engine is also configured to fetch (e.g., from the memory or another memory) a second data block (e.g., a second cache line) that includes data at the second memory address.

Abstract: A technique for performing data prefetching using multi-level indirect data prefetching includes determining a first memory address of a pointer associated with a data prefetch instruction. Content that is included in a first data block (e.g., a first cache line of a memory) at the first memory address is then fetched. A second memory address is then determined based on the content at the first memory address. Content that is included in a second data block (e.g., a second cache line) at the second memory address is then fetched (e.g., from the memory or another memory). A third memory address is then determined based on the content at the second memory address. Finally, a third data block (e.g., a third cache line) that includes another pointer or data at the third memory address is fetched (e.g., from the memory or the another memory).

Abstract: A technique for performing data prefetching using indirect addressing includes determining a first memory address of a pointer associated with a data prefetch instruction. Content, that is included in a first data block (e.g., a first cache line) of a memory, at the first memory address is then fetched. An offset is then added to the content of the memory at the first memory address to provide a first offset memory address. A second memory address is then determined based on the first offset memory address. A second data block (e.g., a second cache line) that includes data at the second memory address is then fetched (e.g., from the memory or another memory). A data prefetch instruction may be indicated by a unique operational code (opcode), a unique extended opcode, or a field (including one or more bits) in an instruction.

Abstract: A method, processor, and data processing system for enabling utilization of a single prefetch stream to access data across a memory page boundary. A prefetch engine includes an active streams table in which information for one or more scheduled prefetch streams are stored. The prefetch engine also includes a victim table for storing a previously active stream whose next prefetch crosses a memory page boundary. The scheduling logic issues a prefetch request with a real address to fetch data from the lower level memory. Then, responsive to detecting that the real address of the stream's next sequential prefetch crosses the memory page boundary, the prefetch engine determines when the first prefetch stream can continue across the page boundary of the first memory page (via an effective address comparison). The PE automatically reinserts the first prefetch stream into the active stream table to jump start prefetching across the page boundary.

Abstract: A system is provided for implementing a multi-tiered full-graph interconnect architecture. In order to implement a multi-tiered full-graph interconnect architecture, a plurality of processors are coupled to one another to create a plurality of processor books. The plurality of processor books are coupled together to create a plurality of supernodes. Then, the plurality of supernodes are coupled together to create the multi-tiered full-graph interconnect architecture. Data is then transmitted from one processor to another within the multi-tiered full-graph interconnect architecture based on an addressing scheme that specifies at least a supernode and a processor book associated with a target processor to which the data is to be transmitted.

Abstract: Mechanisms for providing hardware based dynamic load balancing of message passing interface (MPI) tasks are provided. Mechanisms for adjusting the balance of processing workloads of the processors executing tasks of an MPI job are provided so as to minimize wait periods for waiting for all of the processors to call a synchronization operation. Each processor has an associated hardware implemented MPI load balancing controller. The MPI load balancing controller maintains a history that provides a profile of the tasks with regard to their calls to synchronization operations. From this information, it can be determined which processors should have their processing loads lightened and which processors are able to handle additional processing loads without significantly negatively affecting the overall operation of the parallel execution system. As a result, operations may be performed to shift workloads from the slowest processor to one or more of the faster processors.

Abstract: Mechanisms for modifying an operation of one or more processors executing message passing interface (MPI) tasks are provided. Mechanisms for adjusting the balance of processing work loads of the processors are provided so as to minimize wait periods for waiting for all of the processors to call a synchronization operation. Each processor has an associated hardware implemented MPI load balancing controller. The MPI Load balancing controller maintains a history that provides a profile of the tasks with regard to their calls to synchronization operations. From this information, it can be determined which processors should have their processing loads lightened and which processors are able to handle additional processing loads without significantly negatively affecting the overall operation of the parallel execution system. As a result, operations may be performed to shift workloads from the slowest processor to one or more of the faster processors.

Abstract: A mechanism is provided for assigning memory to on-chip cache coherence domains. The mechanism assigns caches within a processing unit to coherence domains. The mechanism then assigns chunks of memory to the coherence domains. The mechanism monitors applications running on cores within the processing unit to identify needs of the applications. The mechanism may then reassign memory chunks to the cache coherence domains based on the needs of the applications running in the coherence domains. When a memory controller receives the cache miss, the memory controller may look up the address in a lookup table that maps memory chunks to cache coherence domains. Snoop requests are sent to caches within the coherence domain. If a cache line is found in a cache within the coherence domain, the cache line is returned to the originating cache by the cache containing the cache line either directly or through the memory controller.

Abstract: A scalable space-optimized and energy-efficient computing system is provided. The computing system comprises a plurality of modular compartments in at least one level of a frame configured in a hexadron configuration. The computing system also comprises an air inlet, an air mixing plenum, and at least one fan. In the computing system the plurality of modular compartments are affixed above the air inlet, the air mixing plenum is affixed above the plurality of modular compartments, and the at least one fan is affixed above the air mixing plenum. When at least one module is inserted into one of the plurality of modular compartments, the module couples to a backplane within the frame.