On Feb 2, 2005, at 12:43 PM, Jonathan Herriott wrote:
> I compiled LAM/MPI with xlf90, and I am running one of my benchmarks.
> When I specify 1 thread, and I check top, the thread is taking up no
> processor time, so it never does anything. When I specify two
> threads, 1 thread does nothing, and the other one works at 100%. Any
> idea on what might be causing this? When configuring LAM, I had to
> specify --with-fc=/opt/ibmcmp/xlf/8.1/xlf90
> --with-memory-manager=none. I had to do the memory manager one
> because I was getting that problem with -u
> _lam_darwin_malloc_linker_hack.
That almost sounds like a spinlock gone awry. Can you attach a
debugger and take a look at what the two threads are doing? Or (since
you're using OSX), run it through Shark and send the output?
LAM doesn't really do anything with threads that would cause that
behavior, but there are some thing that an application can do to the
MPI layer to make LAM eat up an awful lot of CPU time. Seeing where
the app is spending its time should point us in the right direction.
Thanks,
Brian
--
Brian Barrett
LAM/MPI developer and all around nice guy
Have a LAM/MPI day: http://www.lam-mpi.org/
|
