CPC G06F 9/5061 (2013.01) [G06F 9/4411 (2013.01); G06F 9/54 (2013.01)] | 15 Claims |
1. A computer-implemented method of configuring a cluster of data processing accelerators (DPAs) comprising a plurality of DPAs communicatively coupled to a host device, the method comprising:
receiving, from a driver of the host device at a DPA in the cluster, configuration instructions specific to the DPA that indicate whether to configure the DPA into, or out of, a sub-group of a plurality of sub-groups of DPAs in the cluster, wherein the driver receives processing requests from a plurality of virtual machines and determines a configuration policy comprising the configuration instructions based on an analysis of the processing requests, and wherein the configuration policy optimizes a throughput of the cluster; and
configuring, by the DPA, the DPA according to the configuration instructions specific to the DPA, including:
in response to the configuration instructions specific to the DPA indicating that the DPA is to join a sub-group of DPAs, generating, by the DPA, a communication link to at least one DPA in the sub-group; and
in response to the configuration instructions specific to the DPA indicating that the DPA is to be removed from the sub-group of DPAs, breaking, by the DPA, communication links between the DPA and all other DPAs in the sub-group, wherein responsive to the breaking, the DPA enters a standby state.
|