Quantifying the Effects of Contention on Parallel File Systems

As we move towards the Exactable era of supercomputing, node-level failures are becoming more common-place, frequent check pointing is currently used to recover from such failures in long-running science applications. While compute performance has steadily improved year-on-year, parallel I/O performance has stalled, meaning check pointing is fast becoming a bottleneck to performance. Using current file systems in the most efficient way possible will alleviate some of these issues and will help prepare developers and system designers for Exactable, unfortunately, many domain-scientists simply submit their jobs with the default file system configuration. In this paper, we analyse previous work on finding optimality on Lustre file systems, demonstrating that by exposing parallelism in the parallel file system, performance can be improved by up to 49x. However, we demonstrate that on systems where many applications are competing for a finite number of object storage targets (OSTs), competing tasks may reduce optimal performance considerably. We show that reducing each job's request for OSTs by 40% decreases performance by only 13%, while increasing the availability and quality of service of the file system. Further, we present a series of metrics designed to analyse and explain the effects of contention on parallel file systems. Finally, we re-evaluate our previous work with the Parallel Log-structured File System (PLFS), comparing it to Lustre at various scales. We show that PLFS may perform better than Lustre in particular configurations, but that at large scale PLFS becomes a bottleneck to performance. We extend the metrics proposed in this paper to explain these performance deficiencies that exist in PLFS, demonstrating that the software creates high levels of self-contention at scale.