Placement optimization for virtualized graphics processing

A system, comprising:

• a first virtual compute instance, wherein a virtual graphics processing unit (GPU) is attached to the first virtual compute instance, wherein the first virtual compute instance is implemented using first central processing unit (CPU) resources and first memory resources of a first physical compute instance in a multi-tenant provider network, wherein the virtual GPU is implemented using a physical GPU accessible to the first physical compute instance over a network, wherein the first virtual compute instance is configured to execute an application, and wherein execution of the application on the first virtual compute instance generates one or more GPU instructions that are sent to the virtual GPU over the network and executed using the virtual GPU;
• a second virtual compute instance, wherein the second virtual compute instance is implemented using second CPU resources and second memory resources of a second physical compute instance in the multi-tenant provider network;
• and
• one or more processors and memory storing program instructions executable by the one or more processors to:
• pause the execution of the application on the first virtual compute instance;
• migrate the application to the second virtual compute instance;
• attach the virtual GPU to the second virtual compute instance;
• and
• resume execution of the application on the second virtual compute instance, wherein the execution of the application on the second virtual compute instance generates one or more additional GPU instructions that are sent to the virtual GPU over the network and executed using the virtual GPU.

A computer-implemented method, comprising:

• pausing an application on a first virtual compute instance, wherein the first virtual compute instance is implemented using first central processing unit (CPU) resources and first memory resources of a first physical compute instance, wherein a virtual graphics processing unit (GPU) is attached to the first virtual compute instance, and wherein the virtual GPU is implemented using a physical GPU accessible to the first physical compute instance over a network;
• attaching the virtual GPU to a second virtual compute instance, wherein the second virtual compute instance is implemented using second CPU resources and second memory resources of a second physical compute instance;
• and
• resuming execution of the application on the second virtual compute instance, wherein the execution of the application on the second virtual compute instance generates one or more GPU instructions that are sent to the virtual GPU over the network and executed using the virtual GPU.

One or more non-transitory computer-readable storage media storing program instructions that, when executed on or across one or more processors, perform:

• executing an application on a first virtual compute instance, wherein the first virtual compute instance is implemented using first central processing unit (CPU) resources and first memory resources of a first physical compute instance in a multi-tenant provider network, wherein a virtual graphics processing unit (GPU) is attached to the first virtual compute instance, wherein the virtual GPU is implemented using a physical GPU accessible to the first physical compute instance over a network, and wherein the execution of the application on the first virtual compute instance generates one or more GPU instructions that are sent to the virtual GPU over the network and executed using the virtual GPU;
• pausing the execution of the application on the first virtual compute instance;
• attaching the virtual GPU to a second virtual compute instance, wherein the second virtual compute instance is implemented using second CPU resources and second memory resources of a second physical compute instance in a multi-tenant provider network;
• and
• resuming execution of the application on the second virtual compute instance, wherein the execution of the application on the second virtual compute instance generates one or more additional GPU instructions that are sent to the virtual GPU over the network and executed using the virtual GPU.