Randomized Parallel Prefetching and Buffer Management

Abstract

There is increasing interest in the use of multiple-disk parallel I/O systems to alleviate the I/O bottleneck. Effective use of I/O parallelism requires careful coordination between data placement, prefetching and caching policies. We address the problems of I/O scheduling and buffer management in a parallel I/O system. Using the standard parallel disk model with D disks and a shared I/O buffer of M blocks, we study the performance of on-line algorithms that use bounded lookahead.

We first discuss algorithms for read-once reference strings. It is known (see [3]) that any deterministic prefetching algorithm with either global M-block n local lookahead, must perform a significantly larger number of I/Os than the optimal off-line algorithm. We discuss several prefetching schemes based on a randomized data placement, and present a simple prefetching algorithm that is shown to perform the minimum (up to constants) expected number of I/Os.

For general read-many reference strings, introduce the concept of write-back whereby blocks are relocated between disks during the course of the computation. We show that any on-line algorithm wilh bounded lookahead using deterministic write-back and buffer management policies must have a competitive ratio of Ω(D). We therefore present a randomized algorithm, RAND-WB, that uses a novel randomized write-back scheme. RAND-WB obtains a competitive ratio of $\Theta \left( {\sqrt D } \right)$, which is the best achievable by any on-line algorithm with only global M-block lookahead.