Characterizing Storage Workloads with Counter Stacks

Existing techniques for identifying working set sizes based on miss ratio curves (MRCs) have large memory overheads which make them impractical for storage workloads. We present a novel data structure, the counter stack, which can produce approximate MRCs while using sublinear space. We show how counter stacks can be checkpointed to produce workload representations that are many orders of magnitude smaller than full traces, and we describe techniques for estimating MRCs of arbitrary workload combinations over arbitrary windows in time. Finally, we show how online analysis using counter stacks can provide valuable insight into live workloads.