A Study of Client-based Caching for Parallel I/O

The trend in parallel computing toward large-scale cluster computers running thousands of cooperating processes per application has led to an I/O bottleneck that has only gotten more severe as the the number of processing cores per CPU has increased. Current parallel file systems are able to provide high bandwidth file access for large contiguous file region accesses; however, applications repeatedly accessing small file regions on unaligned file region boundaries continue to experience poor I/O throughput due to the high overhead associated with accessing parallel file system data. In this dissertation we demonstrate how client-side file data caching can improve parallel file system throughput for applications performing frequent small and unaligned file I/O. We explore the impacts of cache page size and cache capacity using the popular FLASH I/O benchmark and explore a novel cache sharing approach that leverages the trend toward multi-core processors. We also explore a technique we call progressive page caching that represents cache data using dynamic data structures rather than fixed-size pages of file data. Finally, we explore a cache aggregation scheme that leverages the high-level file I/O interfaces provided by the PVFS file system to provide further performance enhancements. In summary, our results indicate that a correctly configured middleware-based file data cache can dramatically improve the performance of I/O workloads dominated by small unaligned file accesses. Further, we demonstrate that a well designed cache can offer stable performance even when the selected cache page granularity is not well matched to the provided workload. Finally, we have shown that high-level file system interfaces can significantly accelerate application performance, and interfaces beyond those currently envisioned by the MPI-IO standard could provide further performance benefits