I/O Optimization and Evaluation for Tertiary Storage Systems

Large-scale parallel scientific applications are generating huge amounts of data that tertiary storage systems emerge as a popular place to hold them. SRB, a uniform interface to various storage systems including tertiary storage systems such as HPSS, UniTree etc., becomes an important and convenient way to access tertiary data across networks in a distributed environment. But SRB is not optimized for I/O performance: one SRB I/O call to storage systems must access a contiguous piece of data like UNIX I/O. For many access patterns, this results in numerous I/O calls which are very expensive. In this paper, we present an Optimization Library (SRB-OL) which is built on top of SRB low level I/O functions and employs various state-of-the-art I/O optimizations that could be found in secondary storage systems such as collective I/O and data sieving etc. We also present a novel optimization scheme: superfile that can efficiently deal with large amounts of small files. We also incorporate a subfile technique and other features in SRB such as container, migrate, stage and purge into our SRB-OL. How to use these optimizations is decided by a Meta-data Management System (MDMS) [8] that resides one level above SRB-OL. The user provides access pattern information/hints through user application to MDMS, and then MDMS uses these hints to choose an optimal I/O approach and passes the decision to SRB-OL. Finally, SRB-OL performs optimized SRB I/O calls to access data residing on tertiary storage systems. This layered design and implementation make it possible to design low level libraries independently and to port the high level layers easily. To give a quantitative view of optimized SRB I/O functions, we propose a performance model based on significant I/O experiments. By using this performance model, we can prove that collective I/O, superfile etc have significant performance improvements. In addition, we present an I/O Performance Evaluator that can estimate I/O cost before the user actually carries out her experiment. This provides the user a lot of benefits for running her application.