Evaluation of checkpoint mechanisms for massively parallel machines

Massively parallel machines typically contain thousands of processor units and therefore are more likely to suffer system breakdown because of component failures. This paper studies efficient diskless checkpointing mechanisms for SIMD massively parallel machines. Three checkpointing schemes: mirror checkpointing, parity checkpointing, and partial parity checkpointing are compared in terms of their checkpoint performance and storage overheads, based on empirical measurements. Mirror checkpointing and parity checkpointing schemes have been successfully implemented and tested on a DECmpp 12000 machine, without hardware or OS modifications. It has been shown that mirror checkpointing is an order of magnitude faster than parity checkpointing, but takes twice as much storage overhead. Partial parity checkpointing, although significantly reduces the storage overhead, could lead to unpredictable execution performance. This paper also examines the detailed storage/performance tradeoffs for partial parity checkpointing through manual instrumentation, and describes the implementation experience from these experiments.