Performance tuning of matrix triple products based on matrix structure
PARA'04 Proceedings of the 7th international conference on Applied Parallel Computing: state of the Art in Scientific Computing
| Hi-index | 0.00 |
A detailed understanding of high-performance computer (HPC) file system read and write (I/O) workloads allows stakeholders to evaluate the effectiveness of the I/O infrastructure and identify bottlenecks and other issues. Always-on, server-side monitoring, like that provided by the Lustre Monitoring Tool, permits a comprehensive and nonintrusive mechanism for capturing details of the I/O workload. The statistical properties of data movement to and from mass storage on an HPC system reveal transaction patterns that connect the server-side observations back to the computer-side jobs that caused them. This paper lays out strategies to characterize such patterns using I/O statistics.