xhpl memory leak

Accelerated Computing HPC Compilers Legacy PGI Compilers
1

Hi all,
just wondering if anybody encounter the following problem:
make xhpl binary with with openmpi and ofed 1.2 latest build.
after couple of runs, primary xhpl process grew from 1.2 gig to 2.7 gig
and eventually system ran out of memory.

I build openmpi with gcc and ofed 1.2 latest build, recompile xhpl with openmpi.gcc
and I do not have the problem

thanks for you help in advance
-Brady

2

RHEL5

pgi 7.0

3

could this have something to do with it ?
-DLAZY_MEM_UNREGISTER

4

Hi bchang,

Unfortunately we don’t have mvapich here to test it. We will keep you posted when we have mvapich available.

Removing -DLAZY_MEM_UNREGISTER from CFLAGS might solve the problem based on mvapich discuss forum. I think that was posted by you, perhaps.

http://mail.cse.ohio-state.edu/pipermail/mvapich-discuss/2006-June/000204.html

Hongyon

5

Yes that is mvapich specific macro. I’m using openMpi and I do not find similar flag in openmpi. when I compiled openmpi using gcc I did not see this behavior. Only when I use PGI compiler 6.2, 7.0. I experience xhpl not freeing memory from previous NB



thanks
-Brady

6

Hi bchang,

Since you are using Openmpi with OFED option. I suspect that OPED is compiled with -DLAZY_MEM_UNREGISTER but don’t know for sure.

Is it possible that you can recompile OFED as discussed in mvapich discuss forum?

Hongyon