open TELEMAC-MASCARET The mathematically superior suite of solvers

Hello,

just trying to test Telemac3D on an Intel MIC (vel Xeon Phi)... It concerns comparing MPI runs on a multicore CPU and on the MIC card, example Wesel.

I set in my config/*.cfg file the configuration for the Intel tools:

incs_parallel: -I/panfs/sw/intel_cluster_studio/CS2013.1.046X/impi/4.1.3.048/include64
libs_partel: /panfs/w2/jaj/telemac/v6p3r2/optionals/metis-5.1.0/build/Linux-x86_64/libmetis/libmetis.a
libs_all: -L/panfs/sw/intel_cluster_studio/CS2013.1.046X/impi/4.1.3.048/lib64 -lmpi -lmpiif -lmpigi

compileTELEMAC.py runs smoothly, but when trying to run the Wesel example, with a "princi" user fortran file to be compiled on the run I get an error message from the compiler

ifort: error #10236: File not found: '/panfs/w2/jaj/telemac/v6p3r2/work/wesel3d/t3d_wesel.cas_2014-05-26-16h25min50s/-L/panfs/sw/intel_cluster_studio/CS2013.1.046X/impi/4.1.3.048/lib64'

because the path to the work directory is appended to the flag setting the path to Intel MPI libraries as given in "libs_all". Maybe there's another way (setting a path for the library search?), but I couldn't manage otherwise than modifying these four lines...

jaj@brutus2:/panfs/w2/jaj/telemac/v6p3r2/scripts/python27 > diff runcode.original.py xxx.py
506,507c506,507
< objCmd = objCmd.replace('',path.realpath(cfg[k]))
< exeCmd = exeCmd.replace('',path.realpath(cfg[k]))
---
> objCmd = objCmd.replace('',path.normpath(cfg[k]))
> exeCmd = exeCmd.replace('',path.normpath(cfg[k]))
892,893c892,893
< objCmd = objCmd.replace('',path.realpath(cfg[k]))
< exeCmd = exeCmd.replace('',path.realpath(cfg[k]))
---
> objCmd = objCmd.replace('',path.normpath(cfg[k]))
> exeCmd = exeCmd.replace('',path.normpath(cfg[k]))

Is there a more elegant possibility?

Best regards,
Jacek

Hello Jacek

Nice to see that you are testing Telemac on the new Intel MIC. This will for sure extand the (parallel) Telemac life for a long while and calm down the GPU fever whose community promizes a dark future for MPI-based parallel codes.

with my best regards
Riadh

Hi,

Thank you for the information this bug is already corrected in the development version. So it will be in the next release.

Cheers,
Yugi

Dear Riadh,

yes I am running Telemac on MIC, whereby the aim is to check and confirm that so written (we know how, don't we?) legacy code from the past is not suitable for this kind of present day (co-)processors. This is important for many reasons, especially because Intel announced it will pursue this direction of processor development. As awaited, I have a significant slowdown of execution on a MIC, 10x (ten times) slower than the tandem of two Xeon 8-core processors.

Best regards,
Jacek

And you know the reason(s) behind this efficiency loss? is it due to the fact that the technology is very new and thus not enough effient? may we hope better efficiency soon? Or shall we change in a fundamental way our legacy codes (such is the case for GPU technology)?

Best regards
Riadh

Dear Riadh,

please note my posts from yesterday in the thread named "Potential vectorization errors in BIEF routines" by Telemac-3D, there are some explanations and hints.

Best regards,
Jacek