Viewing Adaptive Migration Policies for Tiered Storage Systems as a Supervised Learning Problem

Many file migration algorithms rely on simple, unchanging, automated heuristics to make file placement decisions for exclusively hierarchical storage systems. Such approaches cannot adapt to changes in the workload or data center configuration. Systems with manually-tuned policies offer a way to deal with changes but require well-trained administrators, thereby driving up the cost of storage management. We are developing an automated and adaptive file migration system for arbitrarily large, tiered storage systems with general topologies. We propose to use techniques developed in the field of machine learning to develop an adaptive migration policy. We frame this policy as a supervised learning problem where the labeled data is derived from file system traces. We use the information gathered from file system traces to solve a regression problem in which we try to predict user workload. Migration decisions are made based on this predicted workload. We do this by defining a cost function that considers the performance characteristics of each part of the tiered storage system. We discuss the implementation of some initial parts of the system, in particular the trace processor and replaying simulator. We then present some preliminary experiments conducted with these parts.