Sharing reconfigurable computing devices between workloads

摘要

One or more physical RCDs (PRCDs) are shared between one or more workloads in one or more virtual computing environments. Example PRCD sharing operations may include: (1) providing a virtual RCD (VRCD) for one of the workloads, the VRCD being programmed with an IC design representing a hardware implementation of a software hotspot in the workload, (2) allocating one of the PRCDs to the workload by scheduling the programmed VRCD on the PRCD, (3) burning the PRCD with the IC design of the programmed VRCD so that the PRCD becomes a programmed PRCD that is capable of implementing the workload's hotspot in hardware, and (4) invoking the programmed VRCD instead of executing the hotspot as software in order to cause the programmed PRCD to implement the hotspot in hardware.

主权项

1. A hypervisor method for sharing one or more physical RCDs (PRCDs) between one or more workloads in one or more virtual computing environments supported by a system of sharable hardware resources that includes one or more CPUs, one or more data storage resources, and said one or more PRCDs, said hypervisor method comprising machine operations that include:
(1) providing a virtual RCD (VRCD) for one of said workloads, said VRCD being programmed with an IC design representing a hardware implementation of a software hotspot in said workload; said programmed VRCD storing said IC design as a hardware implementation file that is native to said one or more PRCDs; said IC design comprising a device-independent register transfer level (RTL) representation of said hotspot that has been synthesized as said IC design; (2) allocating one of said PRCDs to said workload by scheduling said programmed VRCD on said PROD; said PRCDs being grouped into one or more PROD pools each having one or more VRCDs assigned to it, said programmed VRCD only being scheduled on PRCDs in a particular PROD pool to which it is assigned; said programmed VRCD being scheduled on one of said PRCDs on an exclusive basis if the number of VRCDs assigned to said particular PROD pool equals the number of PRCDs therein; said programmed VRCD being scheduled with one or more other programmed VRCDs on a shared basis on one of said PRCDs using a dynamic scheduling policy if the number of VRCDs assigned to said particular PROD pool is greater than the number of PRCDs therein; and (3) burning said PROD with said IC design of said programmed VRCD so that said PROD becomes a programmed PROD that is capable of implementing said workload's hotspot in hardware.