Development of efficient VCPU pinning mechanism in Xen

Abstract

Most of the schedulers in the virtualization platforms provide throughput-oriented policy. However, such throughput-oriented policy has disadvantage in performance for multi-threaded applications which require synchronization among the threads or CPU-intensive programs. To solve this problem, we propose dynamic VCPU pinning mechanism that fixes VCPU to specific physical CPU considering the program characteristics. We demonstrate that the proposed method guarantees improved performance compared to the default VCPU assigning mechanism of credit scheduler in Xen hypervisor.