Rescheduling workload in a hybrid computing environment

A system for managing workloads in a high-throughput computing environment for a hybrid processing system, the system comprising:

• a memory;
• a processor communicatively coupled to the memory;
• and
• a workload manager communicatively coupled to the memory and the processor, the workload manager configured to perform a method comprising:
• retrieving a set of high-throughput computing service level agreements (SLAs), the set of high-throughput computing SLAs being associated with a hybrid processing system comprising at least one server system comprising one or more processors of a first architecture and a set of accelerator systems each comprising a different type of one or more processors of a second architecture that is different from the first architecture, wherein the set of accelerator systems is separate from the at least one server system, and wherein at least one SLA in the set of SLAs is dynamically updatable at least during execution of data parallel workload tasks associated with the at least one SLA, and wherein each SLA in the set of SLAs is a multi-dimensional SLA comprising
• at least one performance throughput SLA,
• at least one performance response time batch window SLA, wherein the performance response time batch window is configured to at least one of shrink and expand during execution of data parallel workload tasks, and
• at least one energy SLA identifying an workload energy consumption threshold, to be satisfied by the hybrid processing system,
• wherein each of the performance throughput SLA, the response time batch window SLA, and the energy SLA is associated with a relative priority between each other;
• dynamically rescheduling, based on the determining, at least a portion of one of the first set of data-parallel workload tasks and the second set of data-parallel workload tasks on a second one of the at least one server system and the set of accelerator systems, the dynamically rescheduling substantially satisfying the one of the performance throughput SLA, the response time batch window SLA, and the energy SLA comprising a highest priority in the relative priority identified to have the highest priority.

A computer program product for managing workloads in a high-throughput computing environment for a hybrid processing system, the computer program product comprising:

• a non-transitory storage medium readable by a processing circuit and storing instructions for execution by the processing circuit for performing a method comprising:
• retrieving a set of high-throughput computing service level agreements (SLAs), the set of high-throughput computing SLAs being associated with a hybrid processing system comprising at least one server system comprising one or more processors of a first architecture and a set of accelerator systems each comprising a different type of one or more processors of a second architecture that is different from the first architecture, wherein the set of accelerator systems is separate from the at least one server system, and wherein at least one SLA in the set of SLAs is dynamically updatable at least during execution of data parallel workload tasks associated with the at least one SLA, and wherein each SLA in the set of SLAs is a multi-dimensional SLA comprising
• at least one performance throughput SLA,
• at least one performance response time batch window SLA, and
• at least one energy SLA,
• to be satisfied by the hybrid processing system,
• wherein each of the performance throughput SLA, the response time batch window SLA, and the energy SLA is associated with a relative priority between each other;
• identifying a first set of data-parallel workload tasks scheduled on the at least one server system and a second set of data-parallel workload tasks scheduled with the set of accelerator systems;
• determining that a set of resources at a first one of the at least one server system and the set of accelerator systems are currently available;
• and
• dynamically rescheduling, based on the determining, at least a portion of one of the first set of data-parallel workload tasks and the second set of data-parallel workload tasks on a second one of the at least one server system and the set of accelerator systems, the dynamically rescheduling substantially satisfying the set of high-throughput computing SLAs.