segfault running BENCHMARK1 on Mac OS X with xlf

Message

amaxwell · #1 Unread post by **amaxwell** » Fri Jan 14, 2005 10:56 pm

I've successfully run the BENCHMARK1 test case on a Linux cluster (Rocks 3.3.0) with the Absoft compiler and MPI, but I'm getting a segfault when I try to run this case on my OS X system (1.8G5 DP, either serial or with OpenMP, IBM xlf 8.1). I've run the damee_4 and upwelling examples successfully on the Mac, so this should be possible.

Stepping through with gdb, the crash appears to occur at CALL lmd_vmix_tile (lmd_vmix.f90, line 70 or so). I'm more familiar with C, so any suggestions on how to debug this would be helpful! This is after time step 0 is written to stdout, BTW.

kate · #2 Unread post by **kate** » Sat Jan 15, 2005 1:52 am

I'm more familiar with xlf on IBM. However, you could be running into memory troubles. Are you linking with -bmaxdata:0x70000000? The default is 256 MB which isn't big enough for BENCHMARK1. How much memory do you have, anyway?

amaxwell · #3 Unread post by **amaxwell** » Sat Jan 15, 2005 3:30 am

kate wrote:I'm more familiar with xlf on IBM. However, you could be running into memory troubles. Are you linking with -bmaxdata:0x70000000? The default is 256 MB which isn't big enough for BENCHMARK1. How much memory do you have, anyway?

I'm not specifying that in the link options, so I'll give it a shot and post the results here. I have 1 GB of RAM, and the ppc/xlf/OpenMP smoked the x86/Absoft/MPI combinations that I've tried; I'm curious to see how it compares with the benchmarks that Rich Signell posted here.

amaxwell · #4 Unread post by **amaxwell** » Mon Jan 17, 2005 4:25 pm

Well, I feel dumb; I'm hitting the stack limit, and the answer was here:
viewtopic.php?t=149

Increasing the number of tiles works for now. Interestingly, -bmaxdata doesn't appear in the xlf user's guide, so perhaps it's specific to the IBM machines?