ifort and Fedora, again

Discussion on computers, ROMS installation and compiling

Moderators: arango, robertson

Post Reply
Message
Author
User avatar
cvl
Posts: 18
Joined: Tue Jun 03, 2003 7:39 pm
Location: NATO Undersea Research Centre
Contact:

ifort and Fedora, again

#1 Post by cvl » Fri Dec 22, 2006 4:50 pm

I'm firing up a ROMS 2.2 build on Fedora core 6 using the latest 64 bit Intel release (IA64 9.1????; NetCDF 3.6.1 built to match).

I've read the previous comments about ulimit, googled my way around the bash and Fedora archives, and have yet to find a proper solution to what I believe might be stacksize issue.

To clarify:
  • 1. Serial BENCHMARK1 runs fine.
    2. OPENMP BENCHMARK1 will only run if ntiles>=16 (I have 2 dual-core 64 bit Xeons, 2GB, and 4MB L2 cache). Otherwise it segfaults.
    3. MPICH2 BENCHMARK1 will run with ntiles>=12. The only reported error is that one of the threads died.
I've tried "ulimit -s unlimited", but suspect it's not propagating to the threads.

I'd like to get the OPENMP runs down to one thread per core... current runs aren't even getting 50% efficiency (16 threads on 4 cores only decreases runtime to .6t_serial; I'd ideally hope for .25t_serial). Anyone know of a compiler, kernel or shell option for any of MPICH2, Linux 26, bash, MPICH2 or ifort that will solve this problem :?:

gfortran worked fine with this configuration, though I've been too absorbed in the ifort builds to see which one is faster. I'll try that after Christmas; though its annoying not to have the ifort option open. :oops:

jpringle
Posts: 86
Joined: Sun Jul 27, 2003 6:49 pm
Location: UNH, USA

#2 Post by jpringle » Sat Dec 23, 2006 1:56 pm

Craig (& others)

I have gone through this on numerous systems. The main trick is figuring out how MPI-whatever runs its various sub-jobs on each system. Does it spawn a csh? a sh? Can you specify? Once you do, you can place the ulimit, etc commands in the appropriate .bashrc .cshrc or whatever.

To debug it, I find it easier to make a little program that reports environmental variables, stack limits, etc, and then I run it with MPIRUN or whatever command works on your machine. That way I can easily tell if whatever change I have made has propogated to the all the jobs started by MPI. If you make this code without any hint of what shell should run it, and have it report the value of $SHELL, you can get an idea of the default behaviour of MPIRUN or MPICH2, or whatever you use.

You can find such a code at

viewtopic.php?t=319&start=0&postdays=0& ... t=jpringle

Cheers,
Jamie

Post Reply