The Research On I/O Requests Scheduling Algorithm Based On Service-aware Machenism

I/O scheduling scheme has a crucial impact on disk performance. The traditionaldisk scheduling schemes improve disk I/O bandwidth utilization by optimizing theseek time, but they don’t take the individual response time into consideration, so thatthey can not satisfy the requirement of real-time, also they can not perceive thedifferent types of QoS requirements from different requests. With the population ofreal-time applications, in order to meet the real-time requests or requests with strongtime-sensitivity, researchers proposed some real-time scheduling schemes, which canmaximize throughput at the same time as meeting deadlines.System needs to meet various QoS requirements proposed by different types ofapplications while processing I/O requests submitted by applications. But differentclasses of applications have different QoS requirements. With computer applicationscomplicated increasingly, it tends to work under the mixed requests mode. Differentapplication requests have different I/O characteristics. Researching and applying I/Ocharacterizations are favourable to optimize or design disk scheduling scheme.Linux kernel has four I/O schedulers. These schedulers have their ownadvantages and disadvantages individually. And they adapt to different applicationsrespectively.The CFQ (Completely Fair Queuing) scheduling scheme becomes defaultscheduler after kernel version2.6.18. After analyzing CFQ schedulers of Linux, wefound that CFQ has weaknesses such as using the amount of processed requests as thestandard of fairness, deploying the RR (Round-Robin) algorithm for queues allocationand over-reliance on priority of request. The weakness of this approach is it will resultin unfairness among queues.In this paper, we propose a service-aware mechanism for the mixed requestsmode. The service-aware mechanism can classify the mixed requests effectivelyaccording to their I/O characteristic of time-sensitivity. Be aware of the weaknessesof CFQ, we present a novel improved CFQ (ICFQ) scheduling algorithm based onservice-aware mechanism. This algorithm can sense requests’ classes effectively viathe service-aware mechanism, aiming to require different scheduling policies fordifferent classes. ICFQ takes disk bandwidth as the standard of fairness, and usesMDRR (Modified Deficit Round Robin) to dispatch queues alternatively. Additionally,ICFQ uses the proportion value as the weight of each queue, which calculated bydividing total requests number of each queue by total requests number of all queues. ICFQ controls the length of time slices by weight and priority.Experiments use IOMeter, a tool used for testing performance. Using throughput,bandwidth, maximum response time and average response time as performanceindicators, we test CFQ and ICFQ respectively. The experimental results show thatICFQ algorithm has improved throughput and bandwidth while maintaining lowerresponse time. Under the mixed requests mode, ICFQ outperforms CFQ and is fairer.