Retrospect: Deterministic Replay of MPI Applications for Interactive Distributed Debugging
Retrospect: Deterministic Replay of MPI Applications for Interactive Distributed Debugging
复制标题
DOI：
10.1007/978-3-540-75416-9_41
复制
发表时间：
2007-09
期刊：
影响因子：
--
通讯作者：
A. Bouteiller;G. Bosilca;J. Dongarra
中科院分区：
文献类型：
--
作者：
A. Bouteiller;G. Bosilca;J. Dongarra
尽管高性能计算被用户热切地采用作为满足对计算能力日益增长的需求的工具，但某些领域仍然缺乏探索。 MPI 范式被认为是过去十年 HPC 基础设施大规模发展的基石。然而，即使在今天，用户也必须面对缺乏能够帮助提高软件堆栈和/或应用程序稳定性的工具。在本文中，我们提出并评估了一种工具，旨在让开发人员能够在并行应用程序的执行时间线上动态地来回移动，从而进一步研究并行应用程序的执行情况。基于不显眼的消息记录机制，强制执行确定性重播，从而提供更简单、更有效的并行软件调试方法。
While high performance computing was eagerly adopted by users as a vehicle for satisfying a growing demand on computational power, some areas are still poorly explored. The MPI paradigm is considered as being the keystone for the large development of the HPC infrastructure over the last decade. However, even today the users have to face the lack of tools able to help increase the stability of the software stack and/or of the applications. In this paper we present and evaluate a tool designed to allow developers to further investigate the execution of parallel applications by enabling them to dynamically move back and forth in the execution timeline of a parallel application. Based on an unobtrusive message logging mechanism, deterministic replay is enforced, leading to a simpler and more efficient way to debug parallel software.