Automatic parameter tuning of hierarchical incremental checkpointing
Automatic parameter tuning of hierarchical incremental checkpointing
复制标题
分层增量检查点自动调参
DOI：
10.1007/978-3-319-17353-5_25
复制
发表时间：
2015
期刊：
影响因子：
--
通讯作者：
and Hiroaki Kobayashi
中科院分区：
文献类型：
--
作者：
Alfian Amrizal;Shoichi Hirasawa;Hiroyuki Takizawa;and Hiroaki Kobayashi
随着未来的HPC系统变得更大，全局文件系统的故障率和检查点成本预计会增加。分层递增CPR是解决这一问题的一种很有前途的方法。它利用本地和全局存储的分层存储系统，并通过在两个连续检查点之间仅写入更新的内存页来执行增量检查点。在本文中，我们回答了一个悬而未决的问题：当检查点开销随时间变化时，如何优化检查点间隔。我们提出了一种运行时检查点间隔自动调整技术来优化分级增量式CPR的效率。评估结果表明，如果能够以适当的检查点间隔来利用存储层次结构，则可以显著提高效率。
As future HPC systems become larger, the failure rates and the cost of checkpointing to the global file system are expected to increase. Hierarchical incremental CPR is a promising approach to solve this problem. It utilizes a hierarchical storage system of local and global storages and performs incremental checkpointing by writing only updated memory pages between two consecutive checkpoints. In this paper, we response to an open question; how to optimize the checkpoint interval when the checkpoint overheads are changing with time as in hierarchical incremental CPR. We propose a runtime checkpoint interval autotuning technique to optimize the efficiency of hierarchical incremental CPR. Evaluation results show that the efficiency can be significantly increased if the storage hierarchy can be exploited with appropriate checkpoint intervals.