[Wien] Segmentation fault in LAPW1 on Opteron / gentoo

[Torsten Andersen]

Dear Olivier,

I experienced the same thing with Opteron when I used the -O5 switch in 
the compiler. Something is buggy in the compiler or the ACML. Compiling 
with -O3 removed the problem for me, although then the speed went down 
on the test_case by a factor of two. However, the speed of the test_case 
is not enough - it also has to work for real systems.

Does anyone have optimized compiler options for the Opteron/PGI that 
also works beyond the test case?

Best regards,
Torsten Andersen.

wiener at arcscluster.caltech.edu wrote:
> Dear Wien2k team,
> 
> We are trying to use Wien2k_04.7 on a dual Opteron cluster running gentoo 
> linux. 
> I am compiling with the PGI-5.1 version of pgf90 and I am linking against 
> ACML-2.0/pgi64_mp . 
> I can get the package to compile with no errors and I can run the 
> test_case both in serial or parallel without problem. I can also run a 
> small case of mine both serial/parallel through a full run_lapw cycle 
> until convergence. However for an intermediate 16 atom supercell witha  
> central impurity, I get a segmentation fault, both in serial or parallel. 
> The memory usage seems to remain low however as I track it with 'top'.
> Does anyone have any suggestion? I appended below the output I get before 
> lapw1 crashes.
> Also, regarding the benchmark posted on the Wien2k site for dual opterons 
> with pgf90/ACML-2.0: was the code compiled in 64bit mode? was it allowing 
> for mp support? did it run fine besides the 'test_case' ?
> 
> Thanks,
> Olivier.
> 
> =================================================
> olivier at strongbad VNi_1e3k $ x lapw1 -p 
> starting parallel lapw1 at Thu Oct 14 18:34:26 PDT 2004
> ->  starting parallel LAPW1 jobs at Thu Oct 14 18:34:26 PDT 2004
> Thu Oct 14 18:34:26 PDT 2004 -> Setting up case VNi_1e3k for parallel 
> execution
> Thu Oct 14 18:34:26 PDT 2004 -> of LAPW1
> Thu Oct 14 18:34:26 PDT 2004 -> 
> running LAPW1 in parallel mode (using .machines)
> Granularity set to 1
> Extrafine unset
> Thu Oct 14 18:34:26 PDT 2004 -> klist:       4
> Thu Oct 14 18:34:26 PDT 2004 -> machines:    node002 node003 node004 
> node005
> Thu Oct 14 18:34:26 PDT 2004 -> procs:       4
> Thu Oct 14 18:34:26 PDT 2004 -> weigh(old):  1 1 1 1
> Thu Oct 14 18:34:26 PDT 2004 -> sumw:        4
> Thu Oct 14 18:34:26 PDT 2004 -> granularity: 1
> Thu Oct 14 18:34:26 PDT 2004 -> weigh(new):  1 1 1 1
> Thu Oct 14 18:34:26 PDT 2004 -> Splitting VNi_1e3k.klist.tmp into junks
> node002
> node003
> node004
> node005
> .machinetmp222
> 4 number_of_parallel_jobs
> prepare 1 on node002
> Thu Oct 14 18:34:26 PDT 2004 -> Creating klist 1 
> [1] 13562
> prepare 2 on node003
> Thu Oct 14 18:34:27 PDT 2004 -> Creating klist 2 
> [2] 13577
> prepare 3 on node004
> Thu Oct 14 18:34:28 PDT 2004 -> Creating klist 3 
> [3] 13592
> prepare 4 on node005
> Thu Oct 14 18:34:29 PDT 2004 -> Creating klist 4 
> [4] 13607
> 
> real    0m3.912s
> user    0m3.733s
> sys     0m0.095s
> [1]    Done                          ( $remote $machine[$p]  ...
> 
> real    0m3.691s
> user    0m3.565s
> sys     0m0.110s
> [2]    Done                          ( $remote $machine[$p]  ...
> waiting for all processes to complete
> 
> real    0m3.826s
> user    0m3.712s
> sys     0m0.096s
> 
> real    0m4.136s
> user    0m4.010s
> sys     0m0.106s
> [4]    Done                          ( $remote $machine[$p]  ...
> [3]  + Done                          ( $remote $machine[$p]  ...
> Thu Oct 14 18:34:33 PDT 2004 -> all processes done.
> **  LAPW1 crashed!
> 0.070u 0.137s 0:08.43 2.3%      0+0k 0+0io 0pf+0w
> 
> 
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at
> http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
> 

-- 
Dr. Torsten Andersen        TA-web: http://deep.at/myspace/
AG Hübner, Department of Physics, Kaiserslautern University
http://cmt.physik.uni-kl.de    http://www.physik.uni-kl.de/