Xen Virtual Machine Scheduling Optimization Based On Cache-relevace

With continuous updates and progress of processing speed and storage capacity of computer, more and more networking tasks and computing tasks use virtualization to share hardware resources. With supporting isolation and cocurrence of multiple virtual machines (VMs), the opensource software Xen, based on paravirtualize technology, is applied in server market more and more widely.There are three scheduling layers:VM scheduling in Xen monitor, process scheduling in VM and thread scheduling in process. The VM scheduling is in the bottom of the whole computer, direct impact on performace of whole system. The scheduling algoritm used by default in Xen allocates credit to VMs to manage scheduling resources, deducting it according to runtime. The algorithm adoptes a simple policy of first come first served, overly concerns fairness of sharing processor resources, while ignoring the I/O resources scheduling, even multiple tasks interfere with each other while cocurrence. As a result, not only the whole system resources utilization but also the performace of each VM decrease.Tasks on VM are divided into three categories:computing-intensive, bandwidth-intensive and latency-sensitive. Computing-intensive tasks focu on CPU resources and scheduling switching frequency, and do not care delays in short-term scheduling; bandwidth-intensive tasks concerns about bandwidth; latency-sensitive tasks run in low frequency but requires low latency. Both of bandwidth-intensive and latency-sensitive tasks run quickly each time, small overhead on prosessor and cache. Above all, the improved scheduling algorithm based on cache-relevance is proposed. It sorts all the VMs by credit and defines the VM occupying current cache as cache-relevant task. Xen chooses the VM of biggest credit from runqueue each0.5ms, if not, cache-relevant task is choosed until its credit is finished, and then chooses the VM of biggest credit. So it can minimize latency of latency-sensitive task and cache switching. Credit of all the VMs will be reset to initial value. Driver domain is forbided to be scheduled when forwarding I/O packets, to reduce swithing and improve whole I/O performace. At last, the paper designs six experiments to test cache-relevace scheduling algorithm, proving correntness and feasibility of the idea.