EAGER: A Transactional, Fault-Aware MPI Prototype to Enable MPI-3 Standardization

并行在计算系统(如多核笔记本电脑)和高端超级计算机中随处可见。为了对这些机器编程，消息传递接口(MPI)是可移植应用程序编程器接口的事实标准，自1993年以来一直由工业界、学术界和政府参与者开发。然而，传统上，当硬件发生故障时，基于MPI的并行程序无法继续工作。随着并行系统在规模和速度上的增长，硬件故障正成为它们面临的更大问题。这项工作为基于MPI的程序提供了一种新的大规模容错方案。本研究通过探索应用事务并行编程技术的可能性，改善了容错MPI的现状。其主要思想是探索基于MPI的并行应用程序如何从事务中获益以实现容错。提出的方法的成功将改变中间件、应用程序编程人员接口和支持此类并行程序的可伸缩子系统(如输入输出系统)的设计。

Parallelism is everywhere in computing systems (such as multicore laptops), and in high end supercomputers. To program these machines, Message Passing Interface (MPI), which is a de facto standard of portable application programmer interfaces, was being developed by industry, academia, and government participants since 1993. However, MPI-based parallel programs have traditionally failed to continue to work when hardware faults occur. Hardware faults are becoming a bigger problem for parallel systems as they grow in size and speed. This work introduces a new solution to fault tolerance at scale for MPI-based programs. This research advances current state of fault-tolerant MPI by exploring a possibility of applying transactional parallel programming techniques. The main idea is to explore how MPI-based parallel applications benefit from transactions to achieve fault tolerance. The success of the proposed approach will change the design of middleware, application programmer interfaces, and scalable subsystems that support such parallel programs (such as input-output systems).