Skip to Main Content
We propose a novel framework for evaluating techniques for power optimization in storage. Given an arbitrary trace of disk requests, we split it into short time intervals, extract a set of simple statistics for each interval, and apply an analytical model to those statistics to obtain accurate information regarding the performance and energy characteristics of the system for that workload. The key abstraction used in our analytical model is the run-length - a single sequential run of requests at the disk level. Using this abstraction, the model is able to account for arbitrary interactions of random and sequential I/Os in the context of a RAID array, and obtain accurate results with less effort than a detailed individual request-level simulation. Various layout and migration policies aimed at power conservation can be easily expressed as transformations on this set of statistics for each time interval. We demonstrate the efficacy of our framework by using it to evaluate PARAID, a recently proposed technique for power optimization in storage arrays. We show that the performance and power predicted by the model under the migration and layout policies of PARAID accurately match the results of a detailed simulation of the system. The analytic model allows us to identify key parameters that affect PARAID performance, and propose an enhancement to the layout of data in PARAID which we show to perform superior to the original technique. We use both the analytic model and detailed simulations to illustrate the benefit of our new layout. This also demonstrates the significant simplicity of evaluating a new technique by applying a high-level model to the extracted trace statistics, compared to the current alternative of either implementing the new technique or simulating it at the level of individual requests.