Phase-detection Based Time-slice Scheduling For Parallel Applications In Virtualized Environment

With the prevelance of virtualization,more and more academic and commercial users explore to run the tightly-coupled parallel applications in cloud platforms.In order to improve the resource utilization,multiple independent virtual machines in cloud platforms are usually consolidated into a single physical machine.This resource over-commitment brings huge synchronization latency to the parallel applications running in virtualized environment,thus severely impact the performance of parallel applications.After analyzing the performance of different types of parallel application in virtualized environment,we find that there are mainly two synchronization overhead of parallel applications running in virtualized environment,they are the kernel-level spinlock latency and the user-level barrier latency.We also find that shortening the time-slice of virtual machines(VMs)can significantly reduce the synchronization overhead of parallel applications.Moreover,a common feature of different types of parallel application at runtime is,there exists an inverse relationship between the change of busy-waiting time in synchronization operations and cache miss rate.This exactly characterizes the runtime behavior of a parallel application in which computation and communication phases are separated from each other.A phase-detection based time-sclice scheduling provide an approach to accelerating the performance of parallel applications in virtualized environment.This method can differentiate computation phases from communication phases of a parallel application at runtime based on the change of cache miss rate.Then it can adaptively adjust the time-slice of virtual CPUs according to different phases,so as to effectively mitigate the synchronization overhead in communication phases while keeping the intensive execution in computation phases unaffected.Experimental results show that,for different types of parallel application running in over-committed environment,the phase-detection based time-slice scheduling method can achieve 1.25-70× performance improvement than credit scheduler of Xen and other representative strategies.