Fourier-assisted machine learning of hard disk drive access time models

Abstract

Predicting access times is a crucial part of predicting hard disk drive performance. Existing approaches use white-box modeling and require intimate knowledge of the internal layout of the drive, which can take months to extract. Automatically learning this behavior is a much more desirable approach, requiring less expert knowledge, fewer assumptions, and less time. Others have created behavioral models of hard disk drive performance, but none have shown low per-request errors. A barrier to machine learning of access times has been the existence of periodic behavior with high, unknown frequencies. We show how hard disk drive access times can be predicted to within 0.83 ms using a neural net after these frequencies are found using Fourier analysis.

Predicting I/O Performance in HPC Using Artificial Neural Networks

“…The idea of the ema-model is to exploit periodic changes in the performance of the storage system. Our analysis of measurements as well as the work of Crume et al [4] indicate that periodic phenomena of access times in sequences of files accesses can be exploited for access time prediction. Figure 2 shows a small series of measurements of sequentially read file accesses.…”

“…Generally, storage systems are modeled in two different ways for access time prediction; either using white-box modeling or black-box modeling [4].…”

“…There are important limitations of white-box modeling, however. For every system an individual model has to be created and a model becomes quite intricate for single hard drives already [4]. To approach the complexity of white-box modeling, Ruemmler and Wilkes analyzed the relevance of different hard drive components for the model deviation to save effort for insignificant parts [11].…”

“…Crume et al developed a method using ANNs which exploits periodic patterns in sequences of file access times [4]. A Fourier analysis is used to determine the most important frequencies, which are then used as input information for the ANNs.…”

“…The different colors correspond to the computed clusters, which are our error classes. Note that due to the random initialization of k-Means, different runs may result in different clusters 4 .…”