A Comparison of Application-Level Fault Tolerance Schemes for Task Pools

Fault tolerance is an important requirement for successful program execution on exascale systems. The common approach, checkpointing, regularly saves a program's state, such that the execution can be restarted after permanent node failures. Checkpointing is often performed on system level, but its deployment on application level can reduce the running time overhead. The drawback of application-level checkpointing is a higher programming expense. It pays off if the checkpointing is applied to reusable patterns. We consider task pools, which exist in many variants. The paper supposes that tasks are generated dynamically and are free of side effects. Further, the final result must be computed from individual task results by reduction. Moreover, the pools must be distributed with private queues, and adopt work stealing. The paper describes and evaluates three application-level fault tolerance schemes for task pools. All use uncoordinated checkpointing and regularly save information in a resilient store. The first scheme (called AllFT) saves descriptors of all open tasks; the second scheme (called IncFT) selectively and incrementally saves only part of them; and the third scheme (called LogFT) logs stealing events and writes checkpoints in parallel to task processing. All schemes have been implemented by extending the Global Load Balancing (GLB) library of the "APGAS for Java" programming system. In experiments with the UTS, NQueens, and BC benchmarks with up to 672 workers, the running time overhead during failure-free execution, compared to a non-resilient version of GLB, was typically below 6%. The recovery cost was negligible, and there was no clear winner among the three schemes. A more detailed performance analysis with synthetic benchmarks revealed that IncFT and LogFT are superior in scenarios with large task descriptors. (C) 2019 Elsevier B.V. All rights reserved.