Adaptive Incremental Checkpointing via Delta Compression for Networked Multicore Systems

摘要

Checkpointing has been widely adopted in support of fault-tolerance and job migration, with checkpoint files preferably kept also at remote storage to withstand unavailability/failures of local nodes in networked systems. Lately, I/O bandwidth to remote storage becomes the bottleneck for checkpointing on a large-scale system. This paper proposes an adaptive incremental checkpointing (AIC), aiming to reduce the checkpointing file size considerably so that its involved overhead is lowered and thus the expected job turnaround time drops. Given production multicore systems are observed to have unused cores often available, we design AIC to make use of separate cores for carrying out multi-level checkpointing with delta compression at desirable points of time adaptively. We develop a new Markov model for predicting the performance of such multi-level concurrent checkpointing, with AIC performance evaluated using six SPEC benchmarks under various system sizes. AIC is observed to lower the normalized expected turnaround time substantially (by up to 47%) when compared to its static counterpart and a recent multi-level checkpointing scheme with fixed checkpoint intervals.