Self-Tuning Optimization on Storage Servers in Parallel File Systems

This paper proposes a framework, which requires keeping tracks of both logical I/O access operations and their corresponding physical access on the storage servers in a parallel file system, to build a self-tuning storage system. Thus, the built self-tuning storage system is able to support dynamical data migrating and data pre-fetching transparently from the view point of clients to boost I/O performance. To this end, we first devised an approach to find out the pairs of logical I/O operations and their associated physical I/O operations. We then employed working set modeling to form I/O access patterns and their corresponding disk access patterns. Finally, with the information about existing access patterns and the similarity analysis of access patterns, it is not difficult to predict the future I/O access operations for possible I/O optimization. Through a series of experiments based on several realistic benchmarks, we show that the newly proposed self-tuning storage system is capable of enabling data migration and data pre-fetching dynamically by using the information about I/O access patterns and the predicted future I/O operations. Therefore, it can improve I/O data throughput significantly for the applications with complicated access patterns; especially, for the applications that require to process multiple-dimensional data, such as medical image processing applications and geographic information system (GIS) applications.