AtFP: Attention-based Failure Predictor for Extreme-scale Computing
Longhao Li, Taieb F. Znati · 2022
Extreme-scale computing is paving the way for unparalleled advances in scientific discovery and innovation. However, as systems scale, their propensity to failure increases significantly, making it difficult for long running applications that span a large number of computing nodes to make forward progress. Achieving high performance in extreme scale environments, while minimizing energy consumption, has emerged as a daunting challenge. Significant advances on how to deal with failure, both physical and logical, have been achieved, with varying degree of success. A key component of fault tolerance relies heavily on the ability of the scheme to predict failure accurately. Varies approaches, including intelligent methods, have been proposed to predict failures. In this paper, we propose an attention-based failure predictor (AtFP), which automatically extracts representative features from the raw event log data to predict failure. The results show that, using the same input and output layers, AtFP outperforms frequently used LSTM methods. The proposed model reduces the F1score by 39% and the training time by 65%.