Modeling Disk Traffic with Bias Methods

Disk traffic modeling is useful in designing effective storage systems. One of the most difficult aspects of modeling disk trace data is understanding the underlying process which generates the trace. This work shows a novel method to model and learn the spatio-temporal locality in the trace generating process by using a conditional distribution based on recent disk accesses observed in the trace. Specifically, we present a class of models where the conditional distribution over the next disk block to access is biased towards recently-accessed disk blocks. Our method is flexible enough to be used to model any temporally-ordered finite sequence of events. We show several variants of this method and show how the method can be used not only to understand the trace-generating process but also to evaluate the performance of a storage system. Our experiments show that our method does a good job of capturing the process generating disk traces.