DYNAMIC ALLOCATION OF PHYSICAL GRAPHICS PROCESSING UNITS TO VIRTUAL MACHINES

摘要

Methods, techniques, and systems for dynamically allocating graphics processing units among virtual machines are provided. Example embodiments provide a dynamic GPU allocation system (“DGAS”), which enables the efficient allocation of physical GPU resources to one or more virtual machines. In one embodiment, the DGAS comprises virtualization logic running on a server computing system that computes GPU benefit factors for the virtual machines on a dynamic basis, and combines the computed GBFs with static priorities to determine a ranked ordering of virtual machines. The available GPU resources are then allocated to some subset of these ranked virtual machines as physical GPU capacity is matched with the requirements of the subset. Physical GPU resources are then allocated to the subset of virtual machines that have the highest promise of GPU utilization.

主权项

1. A computer-implemented method in a server computing system for dynamically allocating physical graphics processing unit (GPU) resources among a plurality of virtual machines to be potentially allocated some portion of the physical GPU resources, the physical GPU resources including processing bandwidth of at least one physical GPU available to the server computing system, comprising:
determining a processing capacity of each physical GPU resource; determining a GPU benefit factor for each virtual machine of the plurality of virtual machines, the GPU benefit factor determined based upon a quantitative assessment of benefit to the virtual machine of using at least a GPU resource; establishing an allocation of the physical GPU resources among the plurality of virtual machines by assigning each physical GPU resource to one or more of the plurality of virtual machines based upon the determined capacity of each physical GPU and the GPU benefit factor of each of the plurality of virtual machines; determining a revised GPU benefit factor for at least one of the plurality of virtual machines based upon runtime behavior of the at least one virtual machine; and dynamically adjusting the physical GPU resource allocation to account for the revised GPU benefit factor determined for the at least one virtual machine.