DDC: Distributed Data Collection Framework for Failure Prediction in Tianhe Supercomputers

Reliability has become an issue to the Tianhe supercomputer series with the scaling of the system. Proactive fault-tolerance based on failure prediction turns into an effective way to improve the system’s fault tolerance ability. Data collection is the basis of the failure prediction which has a great impact on the prediction accuracy, while current data collection methods for failure prediction only got limited data with large overhead. This paper presents DDC data collection framework for failure prediction in Tianhe supercomputers. DDC adopts a distributed data collection architecture which can fully collect the data related to the compute nodes’ health with high efficiency. Through the testing for DDC which ran on TH-1A, the results indicated that DDC had the advantage of low cost and good scalability.