[Developers] Add OpenMP statistics to PUGH
jshalf at lbl.gov
Fri Oct 23 09:39:56 CDT 2009
it is a very bad thing of processes migrate due to NUMA effects.
Data gets pinned to the memory controller that is closest to the
processor that first touched it. Things get very inefficient if the
process migrates to a distant processor.
This cannot be done in a portable way, but if you are on Linux, you
can use the numactl APIs to pin processes and get information about
which process is being occupied.
You might also take a look at JumpShot (inside of MPE) to see how they
do their instrumentation. ftp://ftp.mcs.anl.gov/pub/mpi/mpe/mpe2-1.1.1.tar.gz
It works on BlueGene as well, and profiles any OpenMP implementation
that is built on top of pthreads (as most of them are).
On Oct 23, 2009, at 7:10 AM, Frank Loeffler wrote:
> On Fri, Oct 23, 2009 at 08:56:56AM +0200, Ian Hinder wrote:
>> In addition, we might want to think
>> about adding information about which processors are being used by
>> which threads. You can get this information if you can run "top" on
>> the correct machine, but sometimes this is not possible.
> I am not sure that this can be done in a portable way. I am not even
> sure if this can be done always, let's say on linux. Threads might not
> be bound to one processor, but could migrate from one to another. This
> is not always something bad. I think even mpi processes could migrate
> from one processor to another during a simulation.
> Developers mailing list
> Developers at cactuscode.org
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Developers