Parallel-Disk Buffer Management using Randomizationy

We present an on-line algorithm for prefetching and caching for a multiple-disk parallel I/O system. We introduce the notion of write-back whereby blocks are relocated between disks during the course of the computation. Write-back allows the layout to be altered dynamically to suit access patterns in different parts of the reference string. The algorithm is analyzed using the standard parallel disk model with D parallel disks and a shared I/O buffer of sizeM blocks. We show that any on-line algorithm withM -block lookahead using deterministic write-back and buffer management policies must have a competitive ratio of (D). We therefore present a randomized algorithm, RAND-WB, that uses randomized write-back and attains a competitive ratio of (pD), the best achievable by any on-line algorithm with M -block lookahead. Simulations show that RAND-WB obtains considerable performance benefit even when the data has just a small amount of re-use.