A Case Study of Incremental and Background Hybrid In-Memory Checkpointing

Future exascale computing systems will have high failure rates due to the sheer number of components present i n the system. A classic fault-tolerance technique used in tod ay’s supercomputers is a checkpoint-restart mechanism. Howeve r, traditional hard disk-based checkpointing techniques wil soon hit the scalability wall. Recently, many emerging non-volatile memory technologies , such as Phase-Change RAM (PCRAM), are becoming available and can replace disks with the superior latency and power characteristics. Previous research has demonstrated that taking checkpoints at multiple levels referred to ashybrid checkpointing and employing PCRAM for taking local checkpoints can dramatically reduce checkpoint overhead and has the potential to s cale beyond the exascale. In this work, we develop two prototypes to evaluate hybrid checkpointing. We find that, although global checkpointing is slow, by carefully scheduling checkpointoperations, we can hide its overhead using an extra checkpoint cop y maintained in the local PCRAM of each node. In addition, as local checkpointing gets faster, taking more frequent chec kpoints can help reduce the size of incremental checkpoints. Howeve r, in order to benefit from incremental checkpointing, the checkpoint interval has to be less than 10 seconds.