Under AIX 5L 5.3 TL06-SP1, what would be the 'best; tool for performance monitoring and tuning of POE, OMP and hybrid POE/OMP jobs?
For instance, I would like to see (or collect info equivalent to) space X time diagrams (resources in Y-axis, time in X-axis) with annotations for computation/synchronization or communication per task.
Another issue is thread / task location: I would like to be able to 'trace' the locations of OMP threads / POE tasks as they migrate from processor to processor.
I started looking at pct but it seems slow and clumsy.
Can 'perfwb' provide any of the above info for Pthreaded/OMP code?
This topic has been locked.
2 replies Latest Post - 2008-08-04T18:12:46Z by michael-t
Pinned topic Best tool for performance monitoring / tuning of POE / OMP / hybrid jobs?
Answered question This question has been answered.
Unanswered question This question has not been answered yet.
Updated on 2008-08-04T18:12:46Z at 2008-08-04T18:12:46Z by michael-t
parkes 120000ENPC1 PostACCEPTED ANSWER
Re: Best tool for performance monitoring / tuning of POE / OMP / hybrid jobs?2008-08-04T17:57:02Z in response to michael-tMichael,
We need to understand how your first request differs from what would be offered in a standard trace visualization tool, which can show various events (Y axis) that occurred during an application's execution on a timeline (X axis) that correlates the activities of all the processors involved in the job.
For your second question, about how to trace threads as they get dispatched, the answer is that the AIX system trace is probably the right tool:
Details about the trace facility are available in the General Programming Concepts: Writting and Debugging Programs manual available from http://publib.boulder.ibm.com/infocenter/pseries/v5r3/index.jsp
michael-t 120000PNCE28 PostsACCEPTED ANSWER
Re: Best tool for performance monitoring / tuning of POE / OMP / hybrid jobs?2008-08-04T18:12:46Z in response to parkesYes, what then would a decent trace visualization tool be for AIX and what AIX utility could be the source of these types of traces? Just references to documentation would be fine.
I saw the tprof and other 'profiling' tools which report summaries / process and threads.
When one wants to tune up the execution of multi-threaded code he needs to examine where each thread spends most its time over time. Things get more complex when MPI code executes OMP tasks....
thanks for the reply,