Trace driven analysis of write caching policies for disks

The I/O subsystem in a computer system is becoming the bottleneck as a result of recent dramatic improvements in processor speeds. Disk caches have been effective in closing this gap but the benefit is restricted to the read operations as the write I/Os are usually committed to disk to maintain consistency and to allow for crash recovery. As a result, write I/O traffic is becoming dominant and solutions to alleviate this problem are becoming increasingly important. A simple solution which can easily work with existing tile systems is to use non-volatile disk caches together with a write-behind strategy. In this study, we look at the issues around managing such a cache using a detailed trace driven simulation. Traces from three different commercial sites are used in the analysis of various policies for managing the write cache.We observe that even a simple write-behind policy for the write cache is effective in reducing the total number of writes by over 50%. We further observe that the use of hysteresis in the policy to purge the write cache, with two thresholds, yields substantial improvement over a single threshold scheme. The inclusion of a mechanism to piggyback blocks from the write cache with read miss I/Os further reduces the number of writes to only about 15% of the original total number of write operations. We compare two piggybacking options and also study the impact of varying the write cache size. We briefly looked at the case of a single non-volatile disk cache to estimate the performance impact of statically partitioning the cache for reads and writes.