Mass storage system performance prediction using a trace-driven simulator

Performance prediction of mass storage systems can be difficult because of the complexity of the systems and the interdependence of the components. This difficulty can lead to over or under provisioned systems and can limit the ability to identify software algorithms (such as cache management or request ordering algorithms) that scale well and yield high performance. Moreover, as mass storage systems scale to a petabyte and beyond, the ability to predict performance could become increasingly important since errors in capacity planning could also grow. This paper discusses a trace-driven discrete event simulator that we have developed to aid us in ranking design and configuration alternatives. The simulator reads in a configuration file, ingests a workload and estimates multiple metrics, including average user response times, cache hit ratios and device utilization. Simulated components include tape drives, disk systems and software components.