Contribution Details |
| |
| Name: |
 |
BoF 19: Automatic Node & System Performance Analysis at Scale |
|
| |
| Time: |
 |
Wednesday, June 20, 2012
1:45 PM - 2:30 PM |
|
| |
| Room: |
|
Hall C2.2
CCH - Congress Center Hamburg |
|
| |
| Speakers: |
|
Erik Hagersten, Rogue Wave Software |
|
|
|
Rainer Keller, University of Applied Science Stuttgart |
|
|
|
Bernd Mohr, Jülich Supercomputing Center |
|
| |
| Abstract: |
|
Today's scalable HPC architectures consist of 1000s of nodes, each with 10s of cores and a deep memory hierarchy connecting the cores to a non-uniform node memory. Scalable HPC applications need to take advantage of these architectural features, or risk loosing an order of magnitude in efficiency. Tuning applications in such a way requires time, knowledge and skills. In this session we will discuss two productivity tools that automatically analyze the behavior of an application and suggest the necessary performance optimizations to the programmer. Scalasca (of Jülich Supercomputer Centre) performs its analysis mostly at the system level and detects deficiencies in the way nodes communicate with each other, while ThreadSpotter (of Rogue Wave Software) performs its analysis at the node level and detects inefficient memory accesses with thread-granularity. Real-world case-studies will be used to show how these tools are used today. We would also like to get feedback on future plans to make these tools more interoperable, as well as discuss new features needed towards Exascale. Part of this work is currently carried out in the two European projects HOPSA and H4H. |
|
|
|
|
|
