The Study Of Supercomputer Failures Prediction Based On Log Analysis

Autonomic computing aims to solve the management, maintenance and cost problems in more and more complex computer system. Bringing autonomic to the present computer system can make the system to be self-configuration, self-healing, self- optimizing and self-protection, thereby reduce the cost of management and maintenance of computer system. As the most powerful computer, supercomputer plays an important role in the society, but the stability is becoming a serious concern to the supercomputer system when the application system rapidly grows in size and complexity. And autonomic computing can make the supercomputer system more stability; at the same time bring down the cost of management and maintenance.Failure prediction is of great significance for carrying out the autonomic computing in supercomputer system, thereby improve the self-protection ability of supercomputer system effectively and at the same time make the supercomputer system more stability. Log analysis is an effective way to predict failure. Failure prediction needs a basic framework, including log preprocessing, base predictor and joint predictor. And there are two base predictors according to the characteristic of the log failures—predictor based on time between failures and predictor based on association between failures, and joint predictor based on the base predictor.Taking the log file of BlueGene/L which made by IBM as study object. After the step of preprocessing large amount of log file, time predictor, association predictor and joint predictor based on the base predictor are applied to analyze the log file. The experiment shows that the predictive of joint predictor is better than base predictor because the joint predictor can take the proper base predictor according to the characteristic of the failure.