Reliability hierarchies

Hierarchies of diverse storage levels have been analyzed extensively for their ability to achieve both good performance and low cost. The article argues that we should view hierarchies also as a way to achieve both good reliability and low overhead. After discussing the design of a reliability hierarchy, we suggest two metrics to use in evaluating the overall reliability of a reliability hierarchy: mean time to data loss (MTTDL) and data loss rate. These and other metrics allow researchers to evaluate the reliability/performance tradeoffs quantitatively for new storage devices and hierarchies. We use this framework to evaluate how the Rio file cache affects the mean time to data loss and data loss rate of an existing storage hierarchy. Rio improves MTTDL over a standard delayed-write file cache by an order of magnitude and can be used with a long write-back delay without increasing data loss rate. Rio fills in the reliability gap between ordinary memory and disk by providing a storage level with high performance and intermediate reliability.