Performance comparison of thrashing control policies for concurrent mergesorts with parallel prefetching

Abstract

We study the performance of various run-time thrashing control policies for the merge phase of concurrent mergesorts using parallel prefetching, where initial sorted runs are stored on multiple disks and the final sorted run is written back to another dedicated disk. Parallel prefetching via multiple disks can be attractive in reducing the response times for concurrent mergesorts. However, severe thrashing may develop due to imbalances between input and output rates, thus a large number of prefetched pages in the buffer can be replaced before referenced. We evaluate through detailed simulations three run-time thrashing control policies: (a) disabling prefetching, (b) forcing synchronous writes and (c) lowering the prefetch quantity in addition to forcing synchronous writes. The results show that (1) thrashing resulted from parallel prefetching can severely degrade the system response time; (2) though effective in reducing the degree of thrashing, disabling prefetching may worsen the response time since more synchronous reads are needed; (3) forcing synchronous writes can both reduce thrashing and improve the response time; (4) lowering the prefetch quantity in addition to forcing synchronous writes is most effective in reducing thrashing and improving the response time.