Fine-grained checkpointing in distributed object systems

T. Eirich

english

Abstract: The paper discusses problems of checkpointing in distributed object systems and presents an algorithm suited optimally to their fine-grained structure. Usually, checkpointing algorithms assume nodes or processes as system units. This assumption results in a coarse-grained structure of checkpointing. We will show that this difference in granularity makes usual checkpoint algorithms inadequate. The presented algorithm is adapted to distributed object systems because it requires minimal overhead per object, thread and message. Only the relevant parts of the system are checkpointed and rolled back. Furthermore, checkpointing can be initiated by every thread without prior synchronization. A central coordinating instance is obsolete in the system. The algorithm pertains to the class of coordinated algorithms and therefore freezes parts of the system for a certain period of time. It is shown that freezing can be reduced to a minimum by overlapping ongoing computations with checkpointing.

[Full Paper (ps,http) , 51 kB][Full Paper (pdf) , 65 kB]