GPUCloudSim: an extension of CloudSim for modeling and simulation of GPUs in cloud data centers

Recent years have witnessed an increasing growth in the usage of GPUs in cloud data centers. It is known that conventional virtualization techniques are not directly applicable to GPUs, making it a challenge to effectively take advantage of virtualization benefits. API remoting, full, para and hardware-assisted virtualization methods are adopted to empower VMs with GPU capabilities. With such a diversity in approaches, there is a need for a simulation environment to study the effectiveness of GPU virtualization techniques and evaluate GPU provisioning and scheduling policies in cloud data centers. In order to model and simulate GPU-enabled VMs in cloud data centers, this work proposes and describes a simulator architecture implemented as an extension of CloudSim. The extension eases up conducting experimental studies that otherwise need to be carried out in real cloud infrastructures. It includes models to simulate interference among co-running applications, the overhead of virtualization and power consumption of GPUs. To demonstrate the usefulness of our extension, we study NVIDIA GRID, a hardware-assisted GPU virtualization solution. We show that for situations where the number of VMs outperforms the number of hosts, the first-fit VM placement of VMware Horizon may not be effective. Instead, we suggest a first-fit increasing VM placement algorithm which increases the acceptance rate by 59%, shortens makespan by 25% and saves energy by 21%.