Dynamic Acceleration Of Parallel Applications In Cloud Platforms Based On Adaptive Time-slice Control

With increasing demand of high performance computing(HPC) resources, virtual clusters in cloud have been qualified for running tightly-coupled parallel applications. Such types of applications, however, may still suffer from significant performance degradation, especially due to the common phenomenon of resource over-commitment in cloud. Since various types of applications often coexist in a virtualized cloud system, it is a huge challenge to efficiently accelerate parallel applications and also effectively avoid the negative impact on other non-parallel applications meanwhile.We carefully investigate the performance degradation issue of running parallel applications in virtualized cloud, and find that the granularity of scheduling period(i.e., the time slice of virtual machines) is a critical factor impacting the parallel application's performance. To this end, we devise a system that accelerates parallel application's performance, which provides a dynamic approach based on the adaptive control over time-slice for virtual clusters in cloud environment. The system monitors the spinlock operation in the virtual machine(VM). When a spinlock is in the state of busy waiting, the spinlock latency is collected and sent to the virtual machine monitor(VMM). Once VMM receives the spinlock latency, it will calculate the average spinlock latency of the VM which sent the spinlock latency sample. After that, VMM will compute the time slices of VMs and take the minimum one among them as the uniform time slice. At last, VMM will apply this uniform time slice to the scheduling process of virtual CPUs(VCPUs). Through adaptive control over time slices of VMs, the system can effectively reduce the spinlock latency of VMs, accelerating parallel application's performanceExperiments show that our system can obtain 1.5-10 X performance gain for running parallel applications, than other state-of-the-art related works(including traditional time-slice control scheme like Credit Scheduling of Xen and the well-known accelerating systems like Co-Scheduling and Balance Scheduling), with nearly unaffected impact on the performance of non-parallel applications.