Abstract: Checkpoint/Restart is one of the important approaches for software fault-tolerance. In this paper， the system-level and application-level coordinated Checkpoint/Restart mechanisms for OpenMP programs are presented. The system-level support is introduced for transparency， and it makes shared data saved by all threads together. The semantics-related operations of OpenMP will be separated from and hence independent of low-level systems by the application-level OpenMP checkpoint protocol， which improves portability of the checkpoint system. Based on the presented mechanism， a CCRG OpenMP Checkpoint/Restart system has been implemented. The experiments， such as NPB3.2-OMP， show the overhead of checkpointing and restarting is so limited that the system can be used in large scale programs.

Key words: OpenMP, Checkpoint/Restart, system-level and application-level coordinated

摘要： 检查点/续算是软件容错的重要途径之一。论文描述了一个系统级和应用级混合的OpenMP检查点机制，系统级支持不仅使检查点系统具有了好的透明性，并且使共享数据的保存不再由主线程单独完成，具有良好的数据局部性。应用级OpenMP协议将与OpenMP相关的协议处理独立出来，提高了系统的可移植性。NPB3.2-OMP测试结果表明，检查点和续算所需要的时间开销小，能够满足大规模程序的实际需求。

关键词: OpenMP, 检查点/续算, 系统级和应用级协同