Efficient Checkpointing over Local Area Networks*

Parallel and distributed computing on clusters of workstations is becoming very popular as i t provides a cost effective way for high performance computing. In these systems, the bandwidth of the communication subsystem (Using Ethernet technology) is about an order of magnitude smaller compared to the bandwidth of the storage subsystem. Hence, storing a state in a checkpoint is much more eficient than comparing states over the network. In this paper we present a novel checkpointing approach that enables eficient performance over local area networks. The main idea is that we use two types of checkpoints: comparecheckpoints (comparing the states of the redundant processes to detect faults) and storecheckpoints (where the state is only stored). The store-checkpoints d u c e the rollback needed after a fault is detected, without performing many unnecessary comparisons. As a particular example of this approach we analyzed the DMR checkpointing scheme with store-checkpoints. Our main result i s that the overhead of the execution tame can be significantly reduced when store-checkpoints are introduced. W e have implemented a prototype of the new DMR scheme and run it on workstations connected by a LAN. The experimental results we obtained match the analytical results and show that in some cases the overhead of the DMR checkpointing schemes over LAN's can be improved by as much