An efficient coordinated checkpointing scheme for multicomputers

A new approach for checkpointing multicomputer applications is presented. The checkpointing is initiated and controlled by a checkpoint coordinator, residing either on one of the nodes running the application or on the host processor attached to the multicomputer. A message count is used to determine if any messages are in transit. The proposed strategy is hardware-independent and can be implemented in any multicomputer system irrespective of the architecture, interconnection, and routing strategy. This scheme can be used for FIFO and non-FIFO channels as well as with channels where messages can be lost. Measurement results obtained from our simulations indicate that the proposed strategy outperforms an existing scheme proposed for fixed-path wormhole-routed multicomputer systems. Although the proposed strategy is targeted for high-performance, massively parallel multicomputers, it can also be used in any general-purpose distributed system to improve the checkpointing overhead