Труды Института системного программирования РАН (Oct 2018)
The prospects for virtualization of high performance x64 systems
Abstract
In this paper we explore the prospects of virtualization technologies being applied to x64-based high performance systems. The main reasons for performance overhead when running parallel programs inside multiple virtual machines are discussed. We consider KVM/QEMU and Palacios systems and use HPC Challenge and NASA Advanced Supercomputing Parallel Benchmarks packages as a test suite. All tests are performed using high performance cluster with high-speed Infiniband interconnect.Gathered results show feasibility of applying virtualization to a big class of high performance applications. We use huge memory pages to decrease number of TLB cache misses and NUMA emulation mechanism to provide the description of memory topology of real system to virtual machine. Such optimizations decrease virtualization overhead from 10-60% to 1-5% for most tests from HPC Challenge and NASA Advanced Supercomputing Parallel Benchmarks suites. The main bottlenecks of virtualization systems are decreased memory system performance (critical only for a narrow class of applications), device virtualization overhead, and increased noise level caused by the host operating system and hypervisor. Noise may affect performance and scalability of fine-grained applications (those with frequent communications of small size). While the number of nodes in the system grows, the noise influence substantially increases.
