学术文献库

Taking snapshots of the state of a distributed computation is useful for off-line analysis of the computational state, for later restarting from the saved snapshot, for cloning a copy of the computation, and for migration to a new cluster. The problem is made more difficult when supporting collective operations across processes, such as barrier, reduce operations, scatter and gather, etc. Some processes may have reached the barrier or other collective operation, while other processes wait a long time to reach that same barrier or collective operation. At least two solutions are well-known in the literature: (I) draining in-flight network messages and then freezing the network at checkpoint time; and (ii) adding a barrier prior to the collective operation, and either completing the operation or aborting the barrier if not all processes are present. Both solutions suffer important drawbacks. The code in the first solution must be updated whenever one ports to a newer network. The second solution implies additional barrier-related network traffic prior to each collective operation. This work presents a third solution that avoids both drawbacks. There is no additional barrier-related traffic, and the solution is implemented entirely above the network layer. The work is demonstrated in the context of transparent checkpointing of MPI libraries for parallel computation, where each of the first two solutions have already been used in prior systems, and then abandoned due to the aforementioned drawbacks. Experiments demonstrate the low runtime overhead of this new, network-agnostic approach. The approach is also extended to non-blocking, collective operations in order to handle overlapping of computation and communication.