Towards a Storage Backend Optimized for Atomic MPI-I/O for Parallel Scientific Applications

We consider the challenges of building data management systems that are optimized for parallel I/O of scientific applications. One of those challenges focuses on the requirement of high I/O throughput for concurrent data accesses while obeying MPI atomicity semantics. In most existing implementations, the atomicity is often implemented through locking-based schemes, which have proven inefficient. In this proposal, we argue that a novel versioning-based scheme can avoid the need to perform expensive synchronization by using multiple snapshots of the same data, which is much more efficient. We describe our prototype of a versioning-based storage back-end and report on promising experimental results.