Abstract
Parallel I/O is an integral component of modern high performance computing, especially in storing and processing very large datasets, such as the case of seismic imaging, CFD, combustion and weather modeling. The storage hierarchy includes nowadays additional layers, the latest being the usage of SSD-based storage as a Burst Buffer for I/O acceleration. We present an in-depth analysis on how to use Burst Buffer for specific cases and how the internal MPI I/O aggregators operate according to the options that the user provides during his job submission. We analyze the performance of a range of I/O intensive scientific applications, at various scales on a large installation of Lustre parallel file system compared to an SSD-based Burst Buffer. Our results show a performance improvement over Lustre when using Burst Buffer. Moreover, we show results from a data hierarchy library which indicate that the standard I/O approaches are not enough to get the expected performance from this technology. The performance gain on the total execution time of the studied applications is between 1.16 and 3 times compared to Lustre. One of the test cases achieved an impressive I/O throughput of 900 GB/s on Burst Buffer.
| Original language | English (US) |
|---|---|
| Title of host publication | Lecture Notes in Computer Science |
| Publisher | Springer Nature |
| Pages | 701-711 |
| Number of pages | 11 |
| ISBN (Print) | 9783319676296 |
| DOIs | |
| State | Published - Oct 20 2017 |
Bibliographical note
KAUST Repository Item: Exported on 2020-10-01Acknowledgements: For computer time, this research used the resources of the Supercomputing Core Laboratory at King Abdullah University of Science & Technology (KAUST) in Thuwal, Saudi Arabia.