Using ScaLAPACK with large memory problems

[nathan]

Is there a problem when using ScaLAPACK (1.8.0) when the memory for each process is greater than 2 GB?

The code runs okay until the memory per process reaches about 2 GB. For example, when the memory per process is about 2.7 GB (not a tested minimum for crashing), the following error occurs:

*** Process received signal ***
Signal: Segmentation fault (11)
Signal code: Address not mapped (1)
Failing at address: 0x2aaa2aadb580
[ 0] /lib/libpthread.so.0 [0x2add79bd76c0]
[ 1] ../unix/a.out(zgeru_+0x1be) [0x76dd16]
[ 2] ../unix/a.out(pzgeru_+0x5f2) [0x6301da]
[ 3] ../unix/a.out(pzgetf2_+0x80a) [0x5ae0a6]
[ 4] ../unix/a.out(pzgetrf_+0xaa5) [0x526471]

The system and OS are 64 bit. The code and libraries are compiled for 64 bit.

I noticed that Intel claims to have fixed an address error for large problems (> 2GB) in their Cluster MKL. Is this related to the problem I'm experiencing.

Thanks for any help.

[nathan]

After some searching, I found a fix here that I have tested up to 13 GB per process so far.

[Julien Langou]

Hello Nathan, thanks for coming back to the forum. Yes we definetley need to apply
the patch from John Moyard to ScaLAPACK. We'll do it soon. Julien.