SHF: Small: Advanced Compiler Techniques for Meeting Fault Tolerance Needs of HPC Systems

使高性能应用程序对单个节点的故障具有弹性是当今的一项重大挑战。特别是，需要减少与检查点相关的开销，并需要以有效的方式处理静默数据损坏。俄亥俄州立大学的这个项目正在基于新的见解和方法来解决这些问题。提高检查点和重启效率的工作基于以下观察-当前的检查点协议是在分布式计算环境中开发的，没有利用大多数科学并行程序中的关键属性。该项目正在开发静态和动态分析方法，以确定我们所称的“消息意图”，然后可以使用这些方法来实现应用程序级非协调检查点的自动化，而不需要消息日志记录、保持较小的检查点大小和低成本的恢复。此外，正在开发一种独特的软件方法来处理静默数据损坏。这个想法是让控制程序中发生什么计算和/或通信的数据结构具有弹性，只复制它们和它们的存储。这带来了适度的开销，但防止了静默数据损坏对程序稳定性和正确性的最严重影响。这项研究将减少使高性能系统具有弹性的开销，这反过来将提高效率和资源利用率。

Making high performance applications resilient to failure of individual nodes is a major challenge today. Particularly, there is a need for reducing the overheads associated with checkpointing and for dealing with silent data corruption in an effective fashion. This project at Ohio State University is addressing these problems based on new insights and approaches. The work on improving the efficiency of checkpointing and restart is based on the following observation - current checkpointing protocols were developed in context of distributed computing, and do not exploit key properties seen in most scientific parallel programs. This project is developing static and dynamic analysis methods to determine what we refer to as the "message intent", which can then be used to allow automated application-level uncoordinated checkpointing, but without the need for message logging, keeping the checkpoint sizes small, and recovery low-cost. In addition, a distinct software approach for handling silent data corruption is being developed. The idea is to make the data structures that control what computation and/or communication occurs in the program resilient, by only replicating them and their storage. This introduces modest overheads, but guards against the most drastic impact of silent data corruption on program stability and correctness. This research will result in reducing overheads of making high performance systems resilient, which in turn will improve the efficiency and resource utilization.