Generating Representative Synthetic Workloads: An Unsolved Problem

Synthetic disk request traces are convenient and popular workloads for performance evaluation of storage subsystem designs and implementations. This paper develops an approach for validating synthetic disk request generators. Using this approach, commonly-used simplifying assumptions about workload characteristics (e.g., uniformly-distributed starting addresses and Poisson arrivals) are shown to be inappropriate, often leading to inaccurate performance predictions. Also, workload characteristics that require additional study are identified. 1 Introduction Perhaps the largest difficulty facing storage architects and performance analysts is identifying and obtaining workloads with which to evaluate and compare designs. Many groups and organizations have a small set of traces/benchmarks, but few have a large set and fewer still share these all-important tools. The difficulty with obtaining representative workloads is pervasive and many researchers are forced to use whatever is available ...