International Business Machines Corporation

Gregory Howard Bellows

Joaquin Madruga

Barry L Minor

Benjamin Gale Alexander

Ross A Mikosh

Brian H Horton

Brian D Watt

Dean J Burdick

Jason Nathaniel Dale

James E Carey

Charles J Archer

Philip James Sanders

Matthew W Markland

Gordon G Stewart

Jeffrey M Ceason

Joaquín Madruga

Terminating an accelerator application program in a hybrid computing environment that includes a host computer having a host computer architecture and an accelerator having an accelerator architecture, where the host computer and the accelerator are adapted to one another for data communications by a system level message passing module ('SLMPM'), and terminating an accelerator application program in a hybrid computing environment includes receiving, by the SLMPM from a host application executing on the host computer, a request to terminate an accelerator application program executing on the accelerator; terminating, by the SLMPM, execution of the accelerator application program; returning, by the SLMPM to the host application, a signal indicating that execution of the accelerator application program was terminated; and performing, by the SLMPM, a cleanup of the execution environment associated with the terminated accelerator application program.

Terminating an accelerator application program in a hybrid computing environment

A method for efficient dispatch/completion of a work element within a multi-node data processing system. The method comprises: selecting specific processing units from among the processing nodes to complete execution of a work element that has multiple individual work items that may be independently executed by different ones of the processing units; generating an allocated processor unit (APU) bit mask that identifies at least one of the processing units that has been selected; placing the work element in a first entry of a global command queue (GCQ); associating the APU mask with the work element in the GCQ; and responsive to receipt at the GCQ of work requests from each of the multiple processing nodes or the processing units, enabling only the selected specific ones of the processing nodes or the processing units to be able to retrieve work from the work element in the GCQ.

Reducing cross queue synchronization on systems with low memory latency across distributed processing nodes

Disclosed are a method, a system and a computer program product of operating a data processing system that can include or be coupled to multiple processor cores. In one or more embodiments, each of multiple memory objects can be populated with work items and can be associated with attributes that can include information which can be used to describe data of each memory object and/or which can be used to process data of each memory object. The attributes can be used to indicate one or more of a cache policy, a cache size, and a cache line size, among others. In one or more embodiments, the attributes can be used as a history of how each memory object is used. The attributes can be used to indicate cache history statistics (e.g., a hit rate, a miss rate, etc.).

Customizing function behavior based on cache and scheduling parameters of a memory argument

A system and method operable to manage misaligned direct memory access (DMA) data transfers is provided. This method involves determining a delta between N bytes of data to be copied from within a local side buffer (source location) to a remote buffer (destination location). After the delta is determined a tail of the same length is copied to temporary storage. Then the N bytes of data on the local side buffer minus the tail will be shifted to align the N bytes of data to be copied from within the local side buffer to the starting address of the destination location in the remote buffer. The pre-shifted N bytes of data within the local side buffer may be DMA transferred to the remote buffer. The tail transferred to temporary storage may then be copied from temporary storage to the remote buffer.

Managing misaligned DMA addresses

A system and method operable to manage a message queue is provided. This management may involve out-of-order asynchronous heterogeneous remote direct memory access (RDMA) to the message queue. This system includes a pair of processing devices, a primary processing device and an additional processing device, a memory in storage location and a data bus coupled to the processing devices. The processing devices cooperate to process queue data within a shared message queue wherein when an individual processing device successfully accesses queue data the queue data is locked for the exclusive use of the processing device. When the processing device acquires the queue data, the queue data is locked and the queue data acquired by the acquiring processing device includes the queue data for both the primary processing device and additional processing device such that the processing device has all queue data necessary to process the data and return processed queue data.

Managing an out-of-order asynchronous heterogeneous remote direct memory access (RDMA) message queue

A method efficiently dispatches completes a work element within a multi-node, data processing system that has a global command queue (GCQ) and at least one high latency node. The method comprises: at the high latency processor node, work scheduling logic establishing a local command/work queue (LCQ) in which multiple work items for execution by local processing units can be staged prior to execution; a first local processing unit retrieving via a work request a larger chunk size of work than can be completed in a normal work completion execution cycle by the local processing unit; storing the larger chunk size of work retrieved in a local command/work queue (LCQ); enabling the first local processing unit to locally schedule and complete portions of the work stored within the LCQ; and transmitting a next work request to the GCQ only when all the work within the LCQ has been dispatched by the local processing units.

Method to reduce queue synchronization of multiple work items in a system with high memory latency between processing nodes

A system for efficient dispatch/completion of a work element within a multi-node data processing system. The system comprises a processor performing the functions of: selecting specific processing units from among the processing nodes to complete execution of a work element that has multiple individual work items that may be independently executed by different ones of the processing units; generating an allocated processor unit (APU) bit mask that identifies at least one of the processing units that has been selected; placing the work element in a first entry of a global command queue (GCQ); associating the APU mask with the work element in the GCQ; and responsive to receipt at the GCQ of work requests from each of the multiple processing nodes or the processing units, enabling only the selected specific ones of the processing nodes or the processing units to be able to retrieve work from the work element in the GCQ.

System for reducing data transfer latency to a global queue by generating bit mask to identify selected processing nodes/units in multi-node data processing system

A method for managing hardware resources and threads within a data processing system is disclosed. Compilation attributes of a function are collected during and after the compilation of the function. The pre-processing attributes of the function are also collected before the execution of the function. The collected attributes of the function are then analyzed, and a runtime configuration is assigned to the function based of the result of the attribute analysis. The runtime configuration may include, for example, the designation of the function to be executed under either a single-threaded mode or a simultaneous multi-threaded mode. During the execution of the function, real-time attributes of the function are being continuously collected. If necessary, the runtime configuration under which the function is being executed can be changed based on the real-time attributes collected during the execution of the function.

Method for managing hardware resources within a simultaneous multi-threaded processing system

A system efficiently dispatches/completes a work element within a multi-node, data processing system that has a global command queue (GCQ) and at least one high latency node. The system comprises: at the high latency processor node, work scheduling logic establishing a local command/work queue (LCQ) in which multiple work items for execution by local processing units can be staged prior to execution; a first local processing unit retrieving via a work request a larger chunk size of work than can be completed in a normal work completion/execution cycle by the local processing unit; storing the larger chunk size of work retrieved in a local command/work queue (LCQ); enabling the first local processing unit to locally schedule and complete portions of the work stored within the LCQ; and transmitting a next work request to the GCQ only when all the work within the LCQ has been dispatched by the local processing units.

Reducing queue synchronization of multiple work items in a system with high memory latency between processing nodes

A method provides efficient dispatch/completion of an N Dimensional (ND) Range command in a data processing system (DPS). The method comprises: a compiler generating one or more commands from received program instructions; ND Range work processing (WP) logic determining when a command generated by the compiler will be implemented over an ND configuration of operands, where N is greater than one (1); automatically decomposing the ND configuration of operands into a one (1) dimension (1D) work element comprising P sequentially ordered work items that each represent one of the operands; placing the 1D work element within a command queue of the DPS; enabling sequential dispatching of 1D work items in ordered sequence from to one or more processing units; and generating an ND Range output by mapping the 1D work output result to an ND position corresponding to an original location of the operand represented by the 1D work item.

Method to dynamically distribute a multi-dimensional work set across a multi-core system

A system provides efficient dispatch/completion of an N Dimensional (ND) Range command in a data processing system (DPS). The system comprises: a compiler generating one or more commands from received program instructions; ND Range work processing (WP) logic determining when a command generated by the compiler will be implemented over an ND configuration of operands, where N is greater than one (1); automatically decomposing the ND configuration of operands into a one (1) dimension (1D) work element comprising P sequentially ordered work items that each represent one of the operands; placing the 1D work element within a command queue of the DPS; enabling sequential dispatching of 1D work items in ordered sequence from to one or more processing units; and generating an ND Range output by mapping the 1D work output result to an ND position corresponding to an original location of the operand represented by the 1D work item.

Dynamically distribute a multi-dimensional work set across a multi-core system

Provided are techniques for the managing of command queue dependencies and command queue synchronization. Incoming commands are actively tracked through their dependency relationships. Command dependencies may be tracked across multiple lists, including a submission list and a completion list. Each command on the submission list is prepared for processing and ultimately submitted to command processing logic. Command completion processing is performed on each command on the completion list, including by not limited to removing dependencies from pending commands and possibly queuing pending commands for submission to the command processing logic. Also provided as features of a command queue are a standby barrier, an active barrier and a marker. Standby and active barriers are employed to synchronize and track commands through the command queue. Markers are employed to track commands through the command queue.

Synchronizing commands and dependencies in an asynchronous command queue

Disclosed are a method, a system and a computer program product of operating a data processing system that can include or be coupled to multiple processor cores. The multiple processor cores can be coupled to a memory that can include multiple priority queues associated with multiple respective priorities and store multiple work items. Work items stored in the multiple priority queues can be associated with a bit mask which is associated with a respective priority queue and can be routed to respective groups of one or more processors based on the associated bit mask. In one or more embodiments, at least two groups of processor cores can include at least one processor core that is common to both of the at least two groups of processor cores.

Portioning and routing of work in a multiple processor system

Disclosed are a method, a system and a computer program product of operating a data processing system that can include or be coupled to multiple processor cores. In one or more embodiments, an error can be determined while two or more processor cores are processing a first group of two or more work items, and the error can be signaled to an application. The application can determine a state of progress of processing the two or more work items and at least one dependency from the state of progress. In one or more embodiments, a second group of two or more work items that are scheduled for processing can be unscheduled, in response to determining the error. In one or more embodiments, the application can process at least one work item that caused the error, and the second group of two or more work items can be rescheduled for processing.

Growing community of inventors

Austin, TX, United States of America

Gregory Howard Bellows