Research On Parallel I/O Scheduling Policy With Application-Oriented Features

With the rapid growth of storage capacity and CPU processing power, disk bandwidth capacity, however, has not been increased correspondingly, I/O has become a bottleneck impacting computer performance. Parallel file system and parallel I/O become an effective way to solve the disk data processing and mismatching CPU computing power. The I/O request completion time under a parallel file system is depending on the last sub-I/O completion time, the parallel I/O on different data servers request in random order, thus the I/O request completion time was delayed by the last sub-I/O, which greatly wasted CPU processing time.In order to alleviate the I/O bottleneck, this paper proposes a parallel I/O scheduling algorithm featured in application request, the server can identify different application requests, and schedule requests based on the requirements of different applications. In horizontal angle, the I/O request issued by the same application can be performed on different data servers in the same period of time, so that the parallel I/O on the data server can be completed as much as possible in the same period of time, reducing the request waiting time for each application; in the vertical angle, considering the random access feature of I/O size under actual cases, each data server acquire the overall I/O access size, using the small task first algorithm, and time window strategy to ensure fairness, and reduce the average completion time of I/O request, and improve I/O performance of the system.Experimental results show that application-oriented features of the parallel I/O scheduling can not only significantly reduce the average completion time of application requests, but also avoid the hunger, and improve the I/O performance of the system.