[ET Trac] [Einstein Toolkit] #1620: TimerReport doesn't scale very well.

[Einstein Toolkit]

#1620: TimerReport doesn't scale very well.
----------------------+-----------------------------------------------------
  Reporter:  bmundim  |       Owner:                     
      Type:  defect   |      Status:  new                
  Priority:  minor    |   Milestone:                     
 Component:  Other    |     Version:  development version
Resolution:           |    Keywords:                     
----------------------+-----------------------------------------------------

Comment (by bmundim):

 >The overhead in TimerReport surprises me. TimerReport needs to collect
 information from all processes and sort it, which requires a reduction
 operation and string operations. Maybe one of the algorithms used isn't
 scalable -- we should investigate further.

 That surprised me too. As we start using a larger number of cores for our
 simulations, it might be good to take a look at this.

 >The difference in timings comes likely from overhead between the
 respective measurement intervals. The timing prefixed "[0176]" comes from
 the Cactus scheduler, which calls into Carpet, which then has to perform
 certain operations before calling the actual function, which is then
 measured by the other timer (".../thorns/..."). One of these operations is
 synchronization, but there are also others.

 Thanks a lot for this clarification.

 Cheers,
 Bruno.

-- 
Ticket URL: <https://trac.einsteintoolkit.org/ticket/1620#comment:2>
Einstein Toolkit <http://einsteintoolkit.org>
The Einstein Toolkit