A Consensus-Based Fault-Tolerant Event Logger for High Performance Applications

High-performance computing (HPC) systems traditionally employ rollback-recovery techniques to allow faulttolerant executions of parallel applications. Rollback-recovery based on message logging is an attractive strategy that avoids the drawbacks of coordinated checkpointing in systems with low mean-time between failures (MTBF). Most message logging protocols rely on a centralized event logger to store information (i.e., the determinants) to allow the recovery of an application process. This centralized approach, besides the obvious single point of failure problem, represents a bottleneck for the efficiency of message logging protocols. In this work, we present a fault-tolerant distributed event logger based on consensus that outperforms the centralized approach. We implemented the event logger of MPI determinants using Paxos, a prominent consensus algorithm. Our event logger inherits the Paxos properties: safety is guaranteed even if the system is asynchronous and liveness is guaranteed despite processes failures. Experimental results are reported for the performance of the distributed event logger based both on classic Paxos and parallel Paxos applied to AMG (Algebraic MultiGrid) and NAS Parallel Benchmark applications.