Research On Burst-I/O Scheduling Under Data Center Environments

With the increasing popularity of the Internet, the data volume has grown rapidly inboth individuals and business area. Thus how to store such large amounts of data in aconvenient and safe manner becomes a hot research issue in the field of storage. A datacenter characterized by storage space rental and shared storage facilities is a feasiblesolution for data storage in cloud era. A new technical challenge in the data center is tobetter manage storage resources to provide a high quality of service (QoS), which usuallysparks the interest of many researchers.Burst I/O is a common phenomenon within storage workloads, i.e. unexpectedintensive I/O flow in a random period of time. If we deploy a data center according to thehighest demand, it will increase both storage rationing and management cost; thus suchburst I/O phenomenon results in the increase of the storage service requirement.We introduce a workload scheduling framework to solve the burst I/O problem,where the burst I/Os are dynamically decomposed and isolated and the decomposed I/Osare then restructured. First, we survey the related researches on workload schedulingframework, and analyze the impact of burst I/O on performance, and summarize theapproaches to resources assignment used in the data centers.Second, we design a ‘decomposing+restructuring’ scheduling framework. It isshown that such load scheduling framework can efficiently reduce the cost of serverresource distribution under the premise of guaranteed QoS. We design twoload-decomposing schemes and a load-restructuring scheme. All load-scheduling schemesare implemented in a prototype system called Pandaria, and many functional modules areincorporated to the system, including a frame information module, a load schedulingmodule, a thread pool module, a storage load testing module, and so on.Final, we use DiskSim simulator to assess the performance impact caused by burstI/O, and distinguish both pros and cons of the two load-decomposing schemes, and discussthe evaluation issues of storage resources assignment in data center. The comparative testsin Pandaria showed that our scheduling optimization strategy can achieve betterperformance in the burst-I/O case relative to the original load (i.e., unscheduled load). Our scheduling strategy can also reduce storage resource rationing. Therefore, the number ofI/O requests with constraint of low response time increases by a factor of between4.5and9.0.