HySF: A Striped File Assignment Strategy for Parallel File System with Hybrid Storage

Many algorithms have been proposed to statically assign files in parallel I/O systems. Most of them were built on the basis that the underlying storage devices are homogeneous, while the wide use of SSDs makes hybrid storage architecture composed of SSDs and HDDs to be adopted in many parallel I/O systems. In this way, the previous static assignment algorithms could not fulfill the need of this new storage architecture any more. In this paper, we first develop an extensive hybrid parallel I/O system, in which the SSDs and HDDs are laid in the same system level. Then we build the performance model for both SSDs and HDDs in order to find the appropriate roles for them in the parallel I/O system. Based on the performance model, we propose a novel static file assignment algorithm named HySF for the hybrid parallel I/O system targeted to acquire low response time and load balancing. We reach the low response time goal by assigning files to the appropriate respective storage medium and load balancing goal by sorting the files by the access rate. The experiment results show that the HySF algorithm achieves much better performance on mean response time than the other algorithms and presents great flexibility.