PXFS: A persistent storage model for extreme Scale

The continuing technological progress resulted in sustained increase in the number of transistors per chip as well as improved energy efficiency per FLOPS. This spurred a dramatic growth in aggregate computational performance of the largest supercomputing systems, yielding multiple Petascale implementations deployed in various locations over the world. Unfortunately, these advances did not translate to the required extent into accompanying I/O systems, which primarily saw the improvement in cumulative storage sizes required to match the ever expanding volume of scientific data sets, but little more in terms of architecture or effective access latency. Moreover, while new models of computations are formulated to handle the burden of efficiently structuring the parallel computations in anticipation of the arrival of Exascale systems, a meager progress is observed in the area of storage subsystems. New classes of algorithms developed for massively parallel applications, that gracefully handle the challenges of asynchrony, heavily multithreaded distributed codes, and message-driven computation, must be matched by similar advances in I/O methods and algorithms to produce a well performing and balanced supercomputing system. This paper discusses PXFS, a file system model for persistent objects inspired by the ParalleX model of execution that addresses many of these challenges. An early implementation of PXFS utilizing a well known Orange parallel file system as its back-end via asynchronous I/O layer is also described along with the preliminary performance data. The results show perfect scalability and 3× to 20× times speedup of I/O throughput performance comparing to OrangeFS user interface. Also the PXFS module on OrangeFS with 24 clients sees a 5× to 10× times more throughput than NFS.