Two-Level Checkpointing and Verifications for Linear Task Graphs

Fail-stop and silent errors are omnipresent on large-scale platforms. Efficient resilience techniques must accommodate both error sources. To cope with the double challenge, a two-level checkpointing and rollback recovery approach can be used, with additional verifications for silent error detection. A fail-stop error leads to the loss of the whole memory content, hence the obligation to checkpoint on a stable storage (e.g., an external disk). On the contrary, it is possible to use in-memory checkpoints for silent errors, which provide a much smaller checkpointing and recovery overhead. Furthermore, recent detectors offer partial verification mechanisms that are less costly than the guaranteed ones but do not detect all silent errors. In this paper, we show how to combine all of these techniques for HPC applications whose dependency graph forms a linear chain. We present a sophisticated dynamic programming algorithm that returns the optimal solution in polynomial time. Simulation results demonstrate that the combined use of multi-level checkpointing and verifications leads to improved performance compared to the standard single-level checkpointing algorithm.