Research On Computing Performance Optimization In Virtualization

Virtualization technique is increasingly mature and the performance of processor isincreasingly improved, which make it feasible that multiple guest operating systems run on asingle physical machine or even a single processor core. Virtualization technique is prevalentin server consolidation and HPC. In the virtualization environment, Virtual Machine Monitor(i.e., VMM) takeovers most functions of the traditional operating system, running in thehighest privilege level. VMM can create multiple isolated virtual systems which are calledVirtual Machine (i.e., VM), and the operating system running in the VM is called Guest OS.VMM plays a significant role in the resource management of the virtual platform,dynamically mapping the virtual CPUs of virtual machines to physical CPUs according tochosen scheduling policy.Computing performance is an important metric of the computer system. As virtualizationtechnique hasn’t been matured enough, the performance of virtualization platform is worsethan traditional computing platform, mainly reflecting at system performance. It is necessaryto efficiently schedule multiple dedicated virtual machines concurrently running on the signalphysical host. The virtual machines scheduling algorithm in the VMM plays a significant rolein the performance of the virtualization platform. As a result, the emphasis of this research ishow to improve the computing performance of virtualization by optimizing the virtualmachines scheduling algorithm. In detail, based on in-depth analysis of the Credit schedulingalgorithm in Xen virtualization, we propose an algorithm for scheduling CPU-sharing virtualmachines.A traditional operating system solely dominates all of CPUs in the physical machine,while multiple virtual machines can share CPUs or even a processor core. CPU-sharing is acommon phenomenon in virtualization, which means that multiple VMs concurrently runningthe same processor or core. The Xen virtualization system is widely used by industry, whichdefaultly uses Credit scheduler. In Xen VMM, with the increase in the number of theCPU-sharing virtual machines, each VM will experience more delay for accessing CPU, itwill lead to I/O bound applications experiencing more latency of I/O processing. Therefore inthe CPU-sharing virtual machines system, the performance of I/O bound applications is verylow.In order to improve the performance of I/O bound applications in the CPU-sharingvirtual machines system, we introduce a new class of virtual machines, which is calledlatency-sensitive virtual machines. These virtual machines achieve better performance for I/Obound applications，while fairly sharing CPU with other CPU-sharing virtual machines. Wepropose vCredit scheduler. It schedules each latency-sensitive virtual machines more frequently by using smaller micro time slice, while schedules each non latency-sensitivevirtual machines by using default time slice in the Credit. Latency-sensitive virtual machinescan handle I/O events timely, while not violate the fairness of sharing CPU by CPU-sharingvirtual machines. We implement a prototype of vCredit in the Xen VMM. Experiments showthat vCredit scheduler inherits some characteristics of Credit scheduler, such as fairness andload balancing in symmetric multiprocessor systems, at the same time significantly reducesthe RTT of the network packages and the jitter of UDP datagram, thus achieves betterperformance of applications. vCredit shortens the execution times of Intel MPI benchmarkprograms by half or more.