Techniques to improve the scalability of collective checkpointing at large scale

Scientific and data-intensive computing have matured over the last couple of years in all fields of science and industry. Their rapid increase in complexity and scale has prompted ongoing efforts dedicated to reach exascale infrastructure capability by the end of the decade. However, advances in this context are not homogeneous: I/O capabilities in terms of networking and storage are lagging behind computational power and are often considered a major limitation that that persists even at petascale [1]. A particularly difficult challenge in this context are collective I/O access patterns (which we henceforth refer to as collective checkpointing) where all processes simultaneously dump large amounts of related data simultaneously to persistent storage. This pattern is often exhibited by large-scale, bulk-synchronous applications in a variety of circumstances, e.g., when they use checkpoint-restart fault tolerance techniques to save intermediate computational states at regular time intervals [2] or when intermediate, globally synchronized results are needed during the lifetime of the computation (e.g. to understand how a simulation progresses during key phases). Under such circumstances, a decoupled storage system (e.g. a parallel file system such as GPFS [3] or a specialized storage system such as BlobSeer [4]) does not provide sufficient I/O bandwidth to handle the explosion of data sizes: for example, Jones et al. [5] predict dump times in the order of several hours. In order to overcome the I/O bandwidth limitation, one potential solution is to equip the compute nodes with local storage (i.e., HDDs, SSDs, NVMs, etc.) or use I/O forwarding nodes. Using this approach, a large part of the data can be dumped locally, which completely avoids the need to consume and compete for the I/O bandwidth of a decoupled storage system. However, this is not without drawbacks: the local storage devices or I/O forwarding nodes are prone to failures and as such the data they hold is volatile. Thus, a popular approach in practice is to wait until the local dump has finished, then let the application continue while the checkpoints are in turn dumped to a parallel file system in background. Such a straightforward solution can be effective at hiding the overhead incurred to due I/O bandwidth limitations, but this not necessarily the case: it may happen that there is not enough time to fully flush everything to the parallel file system before the next collective checkpoint request is issued. In fact, this a likely scenario with growing scale, as the failure rate increases, which introduces the need to checkpoint at smaller intervals in order to compensate for this effect. Furthermore, a smaller checkpoint interval also means local dumps are frequent and as such their overhead becomes significant itself.