A prefetching prototype for the parallel file system on the paragon

The initial performance results for the file systems demonstrate that the file system performance is scalable. The access bandwidth seen by the user when using prefetching is also scalable, and given a reasonable overlap between computation and I/O, the benefits from the system prefetching can be very significant. Given that in the normal mode of operation (without prefetching), the data is directly transferred into user's buffer, while in the system level prefetching the data is buffered, performance with prefetching is comparable when there is no overlap of I/O with computation. We also compared the performance of system level prefetching with the explicit prefetching at the user level where the knowledge of what and when to prefetch is known. We observed that the system level prefetching performance of comparable to that of user level prefetching despite an extra level of buffering in the former. That demonstrates that the prefetching prototype's performance is not much below the level expected in the best case. In any such implementation in a large-scale software, there are a large number of parameters that can be studied and need to be evaluated. As a part of the future work, we plan to evaluate the performance of prefetching on much larger systems and study the performance for a greater variety of workloads and access patterns. Furthermore, we plan to implement prefetching in other file I/O modes.