Extending a cluster SSI OS for transparently checkpointing message-passing parallel applications

摘要

Nowadays, clusters are widely used to execute scientific applications. These applications are often message-passing parallel applications with long execution times. Since the number of nodes in clusters is growing, faults are more frequent. Thus the application execution time may be greater than the mean time before failure (MTBF) of the cluster. To avoid restarting application from the beginning, it is desirable that cluster systems provide some fault tolerant mechanisms such as checkpoint/restart. An approach to implement efficiently this mechanism is to implement it directly in the application or in the communication library. Another approach is to implement it in an operating system dedicated to clusters. This is more complex but let you checkpoint/restart any message-passing application whatever the communication library. This paper presents basic mechanisms for system initiated checkpoint of message-passing parallel applications running on top of a cluster. Performance results obtained from a prototype implemented in KERRIGHED Single System Image cluster Operating System based on LINUX are presented.