Checkpointing is a typical technique for fault tolerance, whereas its scalability is limited by the overhead of file access. According to the multi-level file system architecture, the cache-style parallel checkpointing was introduced,which translates global coordinated checkpointing into local file operation by out-of-order pipelining of checkpoint flushing opportunity. The overhead of write-back is hidden effectively to increase the performance and the scalability of parallel checkpointing.%检查点机制是高性能并行计算系统中重要的容错手段,随着系统规模的增大,并行检查点的可扩展性受文件访问的制约.针对大规模并行计算系统的多级文件系统结构,提出了cache式并行检查点技术.它将全局同步并行检查点转化为局部文件操作,并利用多处理器结构进行乱序流水线式写回调度,将检查点的写回时机合理分布,从而有效地隐藏了检查点的写回开销,保证了并行检查点文件访问的高性能和高可扩展性.
展开▼