Efficient Checkpointing on MIMD Architectures (thesis)
Report ID: TR-406-93Author: Plank, James S.
Date: 1993-06-00
Pages: 137
Download Formats: |Postscript|
Abstract:
Presented here are efficient algorithms for checkpointing on MIMD architectures. These algorithms have been implemented on two representative machines: a shared-memory multiprocessor, and a message-passing multicomputer. The algorithms and implementations are evaluated according to three speed metrics: checkpoint time, overhead, and latency. Checkpointing is important as a general means of software fault-tolerance. It is also the backbone of certain program control utilities, such as job-swapping, process migration, and playback debugging. We employ several techniques to minimize the invasiveness of the checkpointer on the target program. Such techniques are main memory checkpointing, copy-on-write, buffering, compression, and the elimination of bottlenecks and extra control messages. The major result of this dissertation is that we can implement efficient checkpointing on MIMD architectures, thereby enhancing the usability of such machines.