E级超级计算机故障预测的数据采集方法

面向未来E级超级计算机，提出用于故障预测的数据采集框架，能够全面采集与计算结点故障相关的状态数据。采用自适应多层分组数据汇集方法，有效解决随着系统规模增长数据汇集过程开销过大的问题。在TH-1A超级计算机上的实现和测试表明，该数据采集框架具有开销小、扩展性好的优点，能够满足未来大规模系统故障预测数据采集的需求。
Aimed at an exascale supercomputer, an FPDC (failure prediction data collection framework) was introduced to fully collect the data related to the state of compute nodes’ health. An adaptive multi-layer data aggregation method was presented for data aggregation with less overhead. Extensive experiments, by implementing FPDC on TH-1A，indicate that the FPDC has the advantage of high efficiency and good scalability.