Enhance Virtualized HPC System Based on I/O Behavior Perception and Asymmetric Scheduling

In virtualized HPC system such as virtual cluster and science cloud, CPU-intensive jobs are always companied by high-intensive I/O operations since different computing nodes perform periodic inter-VM communication to transfer data or synchronize computing result. Since traditional VMM schedulers cannot handle the scheduling scenario with mixed workloads efficiently, inter-VM communication always suffers serious performance regression from scheduling competition and then reduces the performance of entire virtualized HPC system. In order to address this issue, this paper proposes an asymmetric scheduling model based on I/O behavior perception. In this mode, we schedule I/O and computing jobs under isolated cpu subsets to erase their performance interference while optimizing the inter-VM communication through short period round robin scheduling. At the same time, we characterize the runtime I/O behavior of applications at fine temporal granularity and predict their I/O load state using specific online predictor. We will replan the scheduling scheme dynamically through migrating VMs across different cpu subsets if we predict a coming I/O intensity variation and decide the system performance could benefit from this scheduling adjustment. We build a prototype based on Xen-4.1.0 virtual and preliminary test results demonstrate that our approach could efficiently promote the performance of inter-VM communication under virtualized HPC environment while reducing the computing performance degradation caused by I/O-prior scheduling.