Analytic treatment of the reliability and performance of mirrored disk subsystems

An analytic model is developed for predicting the performance and reliability of mirrored disk subsystems. The model includes Markovian dependencies in the request stream, read and write traffic, unit failures, and individual request failures necessitating retries. A workload model is synthesized from measurements and used as example input. Statistical tests verify the Markovian nature of the request stream. The mirrored system is seen to offer substantial enhancement to both system reliability and performance. Server-selection, i.e., which drive to use to serve a read request, is also considered. The standard mirror policy, `serve closest', is seen to be sub-optimal. A mechanism for optimal selection is suggested.