This title appears in the Scientific Report :
2017
Please use the identifier:
http://hdl.handle.net/2128/16589 in citations.
On the Impact of Asynchronous I/O on the performance of the Cube re-mapper at High Performance Computing Scale
On the Impact of Asynchronous I/O on the performance of the Cube re-mapper at High Performance Computing Scale
The high performance computing (HPC) ecosystem is, by design, obsessed with performance optimization. Developing an HPC-specific application requires the proper performance profiling and analysing tools. The high number of compute cores and the complexity of an HPC platform lead these utilities to g...
Saved in:
Personal Name(s): | Sid Lakhdar, Riyane Yacine (Corresponding author) |
---|---|
Contributing Institute: |
Jülich Supercomputing Center; JSC |
Imprint: |
2017
|
Physical Description: |
viii, 45 |
Dissertation Note: |
Masterarbeit, University Grenoble Alpes, 2017 |
Document Type: |
Master Thesis |
Research Program: |
Computational Science and Mathematical Methods |
Link: |
OpenAccess OpenAccess |
Publikationsportal JuSER |
The high performance computing (HPC) ecosystem is, by design, obsessed with performance optimization. Developing an HPC-specific application requires the proper performance profiling and analysing tools. The high number of compute cores and the complexity of an HPC platform lead these utilities to generate and deal with very large performance-trace files. In this context, we have considered enhancing the I/O-access of the Cube re-mapper, a state-of-the-art trace-analysis software for HPC executions. We propose an overlapping-I/O write approach to outperform the time-response of the Cube re-mapper. Thanks to a theoretical study of the general pattern followed by the Cube re-mapper, we show that our method may bring an improvement up to 75% on this pattern. We also show that our custom implementation of the Cube re-mapper allows to reduce significantly the perturbation introduced by overlapping the write threads. Our most enhanced version is thus shown to improve the time-response of the Cube re-mapper by up to 64%. |