Adam Crume, Carlos Maltzahn
04/30/2015 05:25 PM
Computer Science
Access time prediction is important for predicting hard disk drive latencies. Existing approaches use white-box models developed with expert knowledge and can take months to parameterize. Automatically learning behavior is much preferred, requiring less expertise, fewer assumptions, and less time. Black-box modeling has been attempted, but none have demonstrated per-request accuracy. Barriers to machine learning of access times include periodicity with high, unknown frequencies, and a discontinuous access time function. Previously, we showed that recognizing periodicity is crucial for accuracy. Unfortunately, Fourier analysis is expensive and ill-suited for this problem. In this paper, we demonstrate that the model can identify periods itself, bypassing Fourier analysis. We show that neural nets, given specific sinusoids as additional inputs, can predict access times, even across differing track lengths and thus differing periods. This removes a significant barrier and expands the range, although challenges remain. Removing discontinuities via a trigonometric transformation further reduces error.