Virtual Machine Placement solution for VGPU enabled Clouds
Anshuj Garg, Uday Kurkure, Hari Sivaraman, Lan Vu · 2019
Virtual machine placement in cloud set-ups is crucial for the efficient management of resources and meeting customer service level agreements (SLAs). Earlier approaches for VM placement problem did not consider GPU requirement of virtual machines while designing the solution. Newer GPUs by NVIDIA are virtualizable. The multiple virtual machines can share a single physical GPU by time division multiplexing. In our work, we address the problem of VM placement considering virtual machines’ GPU resource requirement. We formulated the vGPU aware VM placement problem as Integer Linear Programming (ILP) problem. The solution of ILP model minimizes the total number of physical GPUs used in the cloud for a given set of GPU access requests. We have also designed two heuristics– VIRD and VIRI, for placement of virtual machines. Our experiment shows that heuristic based algorithms perform better than ILP approach in terms of time taken to reach the solution. Also, our results show that the output of VIRD heuristic is same as the ILP approach.