A Diskless Checkpointing Scheme Based on Vertical Encoding to Lower Fault Tolerance Overhead

摘要

Diskless checkpointing is an effective solution to avoid the I/O bottleneck in disk-based checkpointing for tolerating a small number of node failures in large distributed systems. However, the existing encoding schemes used by diskless checkpointing lead to high communication overheads because of cross node encoding. This negates the advantages of diskless checkpointing, especially in scenarios with a limited network bandwidth. This paper proposes a diskless checkpointing scheme with vertical encoding to address this problem. Vertical encoding eliminates the dependency among nodes and also facilitates a balanced communication. Moreover, an analysis model is developed to obtain an optimal configuration parameter. Experimental results show that the proposed scheme reduces significantly the communication overhead of both checkpointing and fault recovery, with no encoding overhead introduced.