Gvmp: A Multi-Objective Joint Vm and Vgpu Placement Heuristic for Api Remoting-Based Gpu Virtualization and Disaggregation in Cloud Data Centers

The diverse needs of customers drive cloud providers to incorporate more GPU-enabled services. It is known that users barely utilize GPUs. Hence, GPU virtualization techniques are employed to enable GPU sharing and increase resource utilization. Among GPU virtualization techniques, API remoting leverages disaggregation for GPU provisioning, allowing GPUs to be accessed remotely by non-resident VMs. It enables a VM to access any available GPU in the data center which, in turn, provides more flexibility to reach a better placement. It also enables users to choose VM and vGPU instances separately based on computational needs as GPU-enabled VM instances with fixed configurations may not proportionally utilize allocated resources. However, an inefficient VM or vGPU placement may result in poor performance and increase in energy consumption and rejection ratio. There is also a network communication overhead for remotely allocated vGPUs. The main challenge is how to place VMs and vGPUs such that rejection ratio is decreased and the negative performance impacts caused by GPU sharing and remote access are kept at a minimum. In this work, we formally define the joint problem of VM and vGPU placement based on API remoting as a multi-objective ILP model and introduce gVMP as a heuristic to solve it. The proposed method is compared against state-of-the-art heuristics and commercial solutions that are based on API remoting and GRID technologies for the placement of GPU-enabled VMs. We study the effectiveness of the proposed gVMP algorithm for low and high arrival rates in slow and fast networks. Our results show that under different scenarios and in comparison to state-of-the-art policies, our heuristic improves average request duration, efficiency, and SLA by up to 40%, 25% and 29%, respectively.