Conflict Detection Algorithm to Minimize Locking for MPI-IO Atomicity

Abstract

Abstract. Many scientific applications require high-performance concurrent I/O accesses to a file by multiple processes. Those applications rely indirectly on atomic I/O capabilities in order to perform updates to structured datasets, such as those stored in HDF5 format files. Current support for atomicity in MPI-IO is provided by locking around the operations, imposing lock overhead in all situations, even though in many cases these operations are non-overlapping in the file. We propose to isolate non-overlapping accesses from overlapping ones in independent I/O cases, allowing the nonoverlapping ones to proceed without imposing lock overhead. To enable this, we have implemented an efficient conflict detection algorithm in MPI-IO using MPI file views and datatypes. We show that our conflict detection scheme incurs minimal overhead on I/O operations, making it an effective mechanism for avoiding locks when they are not needed.

Efficient Support for MPI-I/O Atomicity Based on Versioning

“…Typically, the whole file is locked for each I/O request and thus concurrent accesses are serialized, which is an obvious source of overhead. To avoid this bottleneck in the case of concurrent non-overlapping accesses to the same shared file, an alternative approach [9] proposes to introduce a mechanism for automatic detection of overlapping I/O and thus avoid locking in this case. However, as acknowledged by the authors of this approach, an unnecessary overhead due to the detection mechanism is then introduced for non-overlapping concurrent I/O.…”

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

“…Typically, the whole file is locked for each I/O request and thus concurrent accesses are serialized, which is an obvious source of overhead. To avoid this bottleneck in the case of concurrent non-overlapping accesses to the same shared file, an alternative approach [9] proposes to introduce a mechanism for automatic detection of non-overlapping I/O and thus avoid locking in this case. However, as acknowledged by the authors of this approach, an unnecessary overhead due to the detection mechanism is then introduced for non-overlapping concurrent I/O.…”