Disk cache design and performance as evaluated in large timessharing and database systems

摘要

We present trace-driven simulations of disk caches using traces from production timesharing and database systems. We compute miss and traffic rations, run lengths ,residency times, memory pollution, etc., with varying block size, prefectch and write algorithms. Prefetch produced limited decreases in miss ratio. Copy-bac, decreased disk writes. Periodic flushng of dirty blocks increased traffic slightly. Write-allocate had little effect. large blocks weren't useful. Limiting cache occurpancy by a single transaction had littel effect. The study is unique in the variety and quality of the data.