[Carpet] Performance measurements

[Ian Hinder]

On Sun, Sep 10, 2006 at 10:18:30PM -0500, Erik Schnetter wrote:
> In order to measure the performance of a mesh refinement simulation  
> on multiple processors, I suggest to use parameters such as the  
> following:
> 
> # Activate standard Cactus timer output
> Cactus::cctk_timer_output = full
> 
> # Activate an additional clock based on MPI
> # (important for runs on multiple processors)
> ActiveThorns = "MPIClock" # This thorn is in AlphaThorns
> 
> # Produce timing information in regular intervals
> # (important since runs may abort early)
> ActiveThorns = "TimerReport" # This thorn is in CactusUtils
> TimerReport::out_every    = 1024
> 
> # Activate Carpet timers
> # (using additional timers for mesh refinement related activities)
> Carpet::output_timers_every = 1024

Are the timing measurements (i.e. grid point updates per processors
per second) and the physical time per hour calculated since the start
of the run, or based on the last few iterations? I would suggest that
the latter would be better, to give an instantaneous measurement of
the speed that did not depend on the history of the run.


> 
> # Activate CarpetLib timers
> # (measuring inter-processor communication)
> CarpetLib::print_timestats_every = 1024
> CarpetLib::print_memstats_every  = 1024
> 
> You should replace the number 1024 with a number of iterations that  
> corresponds to several coarse grid time steps; timing output every  
> ten minutes or so would be reasonable.
> 
> Currently all timer output goes to stdout.  That will be improved soon.
> 
> -erik
> 
> -- 
> Erik Schnetter <schnetter at cct.lsu.edu>
> 
> My email is as private as my paper mail.  I therefore support encrypting
> and signing email messages.  Get my PGP key from www.keyserver.net.
> 
> 
> 



> _______________________________________________
> developers mailing list
> developers at lists.carpetcode.org
> http://lists.carpetcode.org/listinfo/developers


-- 
Ian Hinder
hinder at gravity.psu.edu
http://www.gravity.psu.edu/~hinder