Write-back Cache Timeout Prediction

File systems typically hold newly written data in a cache for a short period of time before writing to disk. Because many files are short-lived, by caching file blocks for longer than their lifetime disk operations are avoided. We evaluate the performance of using an on-line learning algorithm to predict the best timeout to use before committing a data-block to disk. We create a loss function based on costs associated with caching and disk operations. Using a cost-based loss function, it is easy to tune the learning algorithm to perform for different criteria, for example to use less cache space, by adjusting the prices of resources.