Code Profiling using Multiple Profilers on Multiple Machines.

摘要

Use of performance analysis tools to understand bottlenecks of large simulation codes is a common practice that is becoming even more significant with the advent of many-core, multi-core and accelerator-enabled HPC systems. While several sophisticated tools are available, each has assorted strengths and weaknesses. Ease of use is an important factor that varies from tool to tool. In general, the learning curve of a complicated tool will often discourage an eager application developer from investing too much time into learning how to use it. Tool robustness is a quality that is of utmost importance to a user/developer of a large HPC simulation code. The unavoidable increase in cores per chip, chips per node and nodes per system is also demanding scalability. In addition, there are software issues related to language features such as template-heavy C++ codes and interlanguage issues such python interface to C++ codes. Improving robustness will definitely/always be a work in progress for tool developers. The best way to improve quality and robustness is to stress-test the tools with these million-line production codes and keeping the tool developers updated about the outcome of the use of the tools with real-life data-sets. Dealing with large amount of profile data and/or reducing the amount of data collected without compromising quality of information also is a feature that has to be offered by a robust tool. Tool portability among various platforms is another issue, especially for open-source tools. Within the DSRC arena there are several large supercomputing systems, such as AMD-based systems from Cray, Intel-based systems from Cray, and Intel-based systems from IBM. The tools should be usable on all of these systems and should provide consistent, trustworthy performance data regardless of the platform. Finally there is always a request from users for 'insight', instead of mountains of performance data. This issue is much harder to tackle than it seems but is something that could use innovative work, especially as we approach extreme-scale computing.