GPU-remoting latency aware virtual machine migration
Various aspects are disclosed for graphics processing unit (GPU)-remoting latency aware migration. In some aspects, a host executes a GPU-remoting client that includes a GPU workload. GPU-remoting latencies are identified for hosts of a cluster. A destination host is identified based on having a lower GPU-remoting latency than the host currently executing the GPU-remoting client. The GPU-remoting client is migrated from its current host to the destination host.
1. A system comprising:
at least one computing device comprising at least one processor;
at least one memory comprising executable instructions, wherein the instructions, when executed by the at least one processor, cause the at least one computing device to at least:
monitor resources of a plurality of hosts of a datacenter cluster;
identify a host executing a plurality of virtual machines (VM) including a plurality of non-GPU-workload-based virtual machines and a graphics processing unit (GPU)-remoting client virtual machine (VM), the GPU-remoting client VM comprising a GPU workload;
determine that the GPU-remoting client VM is to be migrated in response to an identification that: the GPU-remoting client VM is currently co-located with a GPU remoting server VM, resource utilization of the host is greater than a threshold utilization, and the host lacks migratable non-GPU-remoting virtual machines, wherein the host is identified to lack migratable non-GPU-remoting virtual machines in an instance in which a respective one of the plurality of non-GPU-workload-based virtual machines executed by the host comprise at least one affinity rule with the host or a VM executed using the host;
determine GPU-remoting latencies for at least a subset of the plurality of hosts, a respective GPU-remoting latency corresponding to a host-to-host network latency between: a particular host executing the GPU remoting server VM, and a respective host of the at least the subset of the plurality of hosts, wherein the GPU remoting server VM enables access to a GPU resource using compute unified device architecture call execution;
identify, based on the monitored resources, a destination host comprising a lower GPU-remoting latency than the host; and
migrate the GPU-remoting client VM from the host to the destination host based on the destination host comprising the lower GPU-remoting latency than the host.
2. The system of claim 1 , wherein the destination host comprises a lowest GPU-remoting latency among the at least the subset of the plurality of hosts.
3. The system of claim 1 , wherein the instructions, when executed by the at least one processor, cause the at least one computing device to at least:
identify the at least the subset of the plurality of hosts based on a respective one of the at least the subset of the plurality of hosts comprising sufficient resources for the GPU-remoting client virtual machine.
4. The system of claim 1 , wherein the host is identified to execute the GPU-remoting server virtual machine and the GPU-remoting client virtual machine, and a resource utilization of the host is greater than a threshold utilization.
5. The system of claim 1 , wherein the destination host executes the GPU-remoting server virtual machine, and the destination host comprises sufficient resources for the GPU-remoting client virtual machine.
6. The system of claim 1 , wherein the GPU-remoting client virtual machine accesses the GPU resource based on GPU-remoting Application Programming Interface (API) calls to the GPU-remoting server virtual machine, wherein the GPU resource is local to the GPU-remoting server virtual machine.
7. The system of claim 6 , wherein a respective one of the GPU-remoting API calls comprises data and parameters issued by the GPU workload.
8. A non-transitory computer-readable medium, comprising executable instructions, wherein the instructions, when executed by at least one processor, cause at least one computing device to at least:
monitor resources of a plurality of hosts of a datacenter cluster;
identify a host executing a plurality of virtual machines (VM) including a plurality of non-GPU-workload-based virtual machines and a graphics processing unit (GPU)-remoting client virtual machine (VM), the GPU-remoting client VM comprising a GPU workload;
determine that the GPU-remoting client VM is to be migrated in response to an identification that: the GPU-remoting client VM is currently co-located with a GPU remoting server VM, resource utilization of the host is greater than a threshold utilization, and the host lacks migratable non-GPU-remoting virtual machines, wherein the host is identified to lack migratable non-GPU-remoting virtual machines in an instance in which a respective one of the plurality of non-GPU-workload-based virtual machines executed by the host comprise at least one affinity rule with the host or a VM executed using the host;
determine GPU-remoting latencies for at least a subset of the plurality of hosts, a respective GPU-remoting latency corresponding to a host-to-host network latency between: a particular host executing the GPU remoting server VM, and a respective host of the at least the subset of the plurality of hosts, wherein the GPU remoting server VM enables access to a GPU resource using compute unified device architecture call execution;
identify, based on the monitored resources, a destination host comprising a lower GPU-remoting latency than the host; and
migrate the GPU-remoting client VM from the host to the destination host based on the destination host comprising the lower GPU-remoting latency than the host.
9. The non-transitory computer-readable medium of claim 8 , wherein the destination host comprises a lowest GPU-remoting latency among the at least the subset of the plurality of hosts.
10. The non-transitory computer-readable medium of claim 9 , wherein the instructions, when executed by the at least one processor, cause the at least one computing device to at least:
identify the at least the subset of the plurality of hosts based on a respective one of the at least the subset of the plurality of hosts comprising sufficient resources for the GPU-remoting client virtual machine.
11. The non-transitory computer-readable medium of claim 8 , wherein the host is identified to execute the GPU-remoting server virtual machine and the GPU-remoting client virtual machine, and a resource utilization of the host is greater than a threshold utilization.
12. The non-transitory computer-readable medium of claim 8 , wherein the destination host executes the GPU-remoting server virtual machine, and the destination host comprises sufficient resources for the GPU-remoting client virtual machine.
13. The non-transitory computer-readable medium of claim 8 , wherein the GPU-remoting client virtual machine accesses the GPU resource based on GPU-remoting Application Programming Interface (API) calls to the GPU-remoting server virtual machine, wherein the GPU resource is local to the GPU-remoting server virtual machine.
14. The non-transitory computer-readable medium of claim 13 , wherein a respective one of the GPU-remoting API calls comprises data and parameters issued by the GPU workload.
15. A method performed by instructions executed by at least one processor, the method comprising:
monitoring resources of a plurality of hosts of a datacenter cluster;
identifying a host executing a plurality of virtual machines (VM) including a plurality of non-GPU-workload-based virtual machines and a graphics processing unit (GPU)-remoting client virtual machine (VM), the GPU-remoting client VM comprising a GPU workload;
determining that the GPU-remoting client VM is to be migrated in response to an identification that: the GPU-remoting client VM is currently co-located with a GPU remoting server VM, resource utilization of the host is greater than a threshold utilization, and the host lacks migratable non-GPU-remoting virtual machines, wherein the host is identified to lack migratable non-GPU-remoting virtual machines in an instance in which a respective one of the plurality of non-GPU-workload-based virtual machines executed by the host comprise at least one affinity rule with the host or a VM executed using the host;
determining GPU-remoting latencies for at least a subset of the plurality of hosts, a respective GPU-remoting latency corresponding to a host-to-host network latency between: a particular host executing the GPU remoting server VM, and a respective host of the at least the subset of the plurality of hosts, wherein the GPU remoting server VM enables access to a GPU resource using compute unified device architecture call execution;
identifying, based on the monitored resources, a destination host comprising a lower GPU-remoting latency than the host; and
migrating the GPU-remoting client VM from the host to the destination host based on the destination host comprising the lower GPU-remoting latency than the host.
16. The method of claim 15 , wherein the destination host comprises a lowest GPU-remoting latency among the at least the subset of the plurality of hosts.
17. The method of claim 15 , further comprising:
identifying the at least the subset of the plurality of hosts based on a respective one of the at least the subset of the plurality of hosts comprising sufficient resources for the GPU-remoting client virtual machine.
18. The method of claim 15 , wherein the host is identified to execute the GPU-remoting server virtual machine and the GPU-remoting client virtual machine, and a resource utilization of the host is greater than a threshold utilization.
19. The method of claim 15 , wherein the destination host executes the GPU-remoting server virtual machine, and the destination host comprises sufficient resources for the GPU-remoting client virtual machine.
20. The method of claim 15 , wherein the GPU-remoting client virtual machine accesses the GPU resource based on GPU-remoting Application Programming Interface (API) calls to the GPU-remoting server virtual machine, wherein the GPU resource is local to the GPU-remoting server virtual machine.