Dynamic time slice management based on cpupool in virtualized systems

Abstract

In virtualized systems, virtual time that is experienced by virtual machines (VMs) is physically discontinuous. Even if a VM assumes that it uses physical CPUs exclusively, physical CPUs are shared by other VMs, which results in the virtual time being physically discontinuous. Due to this virtual time discontinuity, lock intensive VMs and I/O intensive VMs are slowed down severely. Although many studies have been conducted to solve these problems, most of them try to solve each phenomenon, not the root cause of the problem, the virtual time discontinuity problem. Recently, Ahn et al. generalized these problems into the virtual time discontinuity problem and suggested an architectural solution. They proposed to shorten the time slice to 1ms to remove the virtual time discontinuity. Additionally, to relieve the negative effects on caches caused by frequent context switches, they proposed to use a context prefetcher and to adopt a context aware cache insertion policy. Although they have solved the virtual time discontinuity problem successfully, they suggested architectural solutions that are not available in current processors. In this paper, we try to solve the virtual time discontinuity problem without additional hardware modification. Using cpupools, which are available in Xen, we try to solve the virtual time discontinuity problem by providing a short time slice cpupool to the VMs that suffer from the virtual time discontinuity problem, while providing a long time slice cpupool to the VMs that benefit more from caches. We decide the characteristic of VMs on-the-fly using decision trees that are constructed from experiments. We constructed the decision trees using many different factors that are closely related to the time slice preference of VMs. We used the pause-loop exiting (PLE) count, the inter-processor interrupt (IPI) count, and the number of packets handled per second. Cpupool is a nice solution for managing time slices in the CPU level. However, using cpupools can cause underutilization of CPUs. Underutilized CPUs cannot be used by the VMs that are located in different cpupools, and this can result in low CPU utilization. To solve this problem, we propose a pseudo-work-conserving mechanism between cpupools. Additionally, we identify and mitigate the cache interference problem that can be caused by solving the virtual time discontinuity problem.