vPHI: Enabling Xeon Phi Capabilities in Virtual Machines

Heterogeneous processing has gained popularity in the high performancecomputing (HPC) area lately and it appears to have a great potential for future data centers. In this regard, accelerators, such as GPUs and Intel Xeon Phi, have already started to play a significant role in HPC systems offering a high degree of parallelism to application developers. Furthermore, hardware virtualization is gaining interest in these domains as well, due to thebenefits it offers, such as server consolidation, costs reduction and resourceutilization. In this context, there is a growing potential for integratingaccelerators into the virtualization stacks of the next generation. In this paper we take a first step towards enabling Intel Xeon Phi capabilitiesin virtual machines. We present vPHI, a framework for efficient virtualizationof Intel Xeon Phi resources. To our knowledge, vPHI is the first approach thatenables Xeon Phi sharing between multiple VMs running on the same physical node. We present vPHI as a proof-of-concept using QEMU-KVM as the hypervisor. vPHI utilizes Intel's SCIF API as the transport layer for transferring data over the PCIe to the accelerator device. Preliminary evaluation has shown promising results with vPHI being able to attain 72% of native execution in terms of peak throughput.