Predicting Performance of Non-contiguous I/O with Machine Learning

Abstract

Data sieving in ROMIO promises to optimize individual non-contiguous I/O. However, making the right choice and parameterizing its buffer size accordingly are non-trivial tasks, since predicting the resulting performance is difficult. Since many performance factors are not taken into account by data sieving, extracting the optimal performance for a given access pattern and system is often not possible. Additionally, in Lustre, settings such as the stripe size and number of servers are tunable, yet again, identifying rules for the data-centre proves challenging indeed.

In this paper, we (1) discuss limitations of data sieving, (2) apply machine learning techniques to build a performance predictor, and (3) learn and extract best practices for the settings from the data. We used decision trees as these models can capture non-linear behavior, are easy to understand and allow for extraction of the rules used. Even though this initial research is based on decision trees, with sparse training data, the algorithm can predict many cases sufficiently. Compared to a standard setting, the decision trees created are able to improve performance significantly and we can derive expert knowledge by extracting rules from the learned tree. Applying the scheme to a set of experimental data improved the average throughput by 25–50 % of the best parametrization’s gain. Additionally, we demonstrate the versatility of this approach by applying it to the porting system of DKRZ’s next generation supercomputer and discuss achievable performance gains.

We want to express our gratitude to the “Deutsches Zentrum für Luft- und Raumfahrt e.V.” as responsible project agency and to the “Bundesministerium für Bildung und Forschung” for the financial support under grant 01 IH 11008 A-C.