Merging and Prioritizing Optimization in Block I/O Scheduling of Disk Storage

Abstract

I/O merging optimization at the block I/O layer of disk storage is widely adopted to reduce I/O response time. But it may result in certain overhead of merging judgment in the case of a large number of concurrent I/O requests accessing disk storage, and place negative effects on the response of small requests. This paper proposes a divide and conquer scheduling scheme at the block layer of I/O stack, to satisfy a large number of concurrent I/O requests with less I/O response time and ensure the fairness of each request response by decreasing the average I/O latency. First, we propose a horizontal visibility graph-based approach to cluster relevant block requests, according to their offsets (i.e., logic block numbers). Next, it carries out the optimization operation of merging consecutive block I/O requests within each cluster, as only these requests in the same cluster are most likely to be issued by a specific application. Then, we have introduced the functionality of merging judgment when performing merging optimization to effectively guarantee the average I/O response time. After that, the merged requests in the queue will be reordered on the basis of their priorities, to purposely cut down the average I/O response time. Finally, the prioritized requests are supposed to be delivered to the disk storage, for being serviced. Through a series of experiments, we show that compared to the benchmark, the newly proposed scheme can not only cut down the I/O response time by more than 18.2%, but also decrease the average I/O response time up to 71.7%.