CSR: Eager: HPC Virtualization with SR-IOV

点击翻译按钮获取中文摘要

The recently introduced Single Root I/O Virtualization (SR-IOV)technique for InfiniBand and High Speed Ethernet provides native I/Ovirtualization capabilities and enables us to provision the internalPCI bus interface between multiple Virtual Machines (VMs). However,achieving near native throughput for HPC applications that use bothpoint-to-point and collective operations on virtualized multi-coresystems with SR-IOV presents a new set of challenges for the designersof high performance middleware, such as MPI. In order to solve thisproblem, this project aims to address the following set of challenges:1) How to redesign MPI communication library to achieve efficientlocality-aware communication and facilitate fair resource sharing onmodern virtualized high performance clusters, with SR-IOV? 2) Cancommunication libraries be designed to deliver the best communicationperformance across different VM subscription policies and networkcommunication modes? 3) What are the the challenges involved indesigning support for advanced features such as, live migration,Quality of Service, and I/O storage virtualization? and 4) What kindof benefits, in terms of performance and scalability, can be achievedby the proposed approach for HPC applications? A synergistic andcomprehensive research plan is proposed to address the abovechallenges for HPC Virtualization on clusters with SR-IOV and studyits impact for a set of HPC applications.