Drive-Thru: Fast, Accurate Evaluation of Storage Power Management

Daniel Peek and Jason Flinn

Abstract

Running traces of realistic user activity is an important step in evaluating storage power management. Unfortunately, existing methodologies that replay traces as fast as possible on a live system cannot be used to evaluate timeout-based power management policies. Other methodologies that slow down replay to preserve the recorded delays between operations are too time-consuming. We propose a hybrid approach, called Drive-Thru, that provides both accuracy and speed of evaluation by separating time-dependent and time-independent activity. We first synchronously replay file system activity on the target platform to create a base trace that captures the semantic relationship between file system activity and storage accesses. We then use the base trace as input to a simulator that can evaluate different disk, network, file cache, and file system power management policies. We use Drive-Thru to study the benefit of several recent proposals to reduce file system energy usage.