I have built my 6.5 beta with gfortran 4.8.2, default compiler settings, on 64 bit Ubuntu 14.04 with OpenBLAS 0.2.11 as BLAS.
The li2h2_tce_ccsd completes during serial execution. The test evaluation fails, but I think that's the fault of the evaluation procedure rather than the actual output. Ignoring the first 6 normal modes, the results are in good agreement with the reference data:
Reference:
7 8 9 10 11 12
Frequency 489.60 573.34 865.88 937.51 1072.22 1146.66
1 0.10486 0.00000 0.00000 0.00000 0.65864 -0.69651
2 0.00000 0.00000 0.64073 -0.65864 0.00000 0.00000
3 0.00000 0.65862 0.00000 0.00000 0.00000 0.00000
4 0.00000 0.00000 -0.11088 0.00000 -0.09461 0.00000
5 0.26398 0.00000 0.00000 0.09461 0.00000 0.03974
6 0.00000 -0.09463 0.00000 0.00000 0.00000 0.00000
7 0.00000 0.00000 0.11088 0.00000 -0.09461 0.00000
8 -0.26398 0.00000 0.00000 0.09461 0.00000 -0.03974
9 0.00000 -0.09463 0.00000 0.00000 0.00000 0.00000
10 -0.10486 0.00000 0.00000 0.00000 0.65864 0.69651
11 0.00000 0.00000 -0.64073 -0.65864 0.00000 0.00000
12 0.00000 0.65862 0.00000 0.00000 0.00000 0.00000
Computed by 6.5 beta:
7 8 9 10 11 12
Frequency 489.60 573.32 865.88 937.51 1072.22 1146.66
1 0.10486 0.00000 -0.00000 -0.00000 -0.65864 -0.69651
2 -0.00000 0.00000 0.64073 -0.65864 0.00000 0.00000
3 0.00000 0.65862 0.00000 0.00000 0.00000 0.00000
4 0.00000 0.00000 -0.11088 0.00000 0.09461 0.00000
5 0.26398 0.00000 0.00000 0.09461 -0.00000 0.03974
6 0.00000 -0.09463 0.00000 0.00000 0.00000 0.00000
7 -0.00000 0.00000 0.11088 -0.00000 0.09461 -0.00000
8 -0.26398 0.00000 0.00000 0.09461 -0.00000 -0.03974
9 0.00000 -0.09463 0.00000 0.00000 0.00000 0.00000
10 -0.10486 0.00000 -0.00000 0.00000 -0.65864 0.69651
11 0.00000 0.00000 -0.64073 -0.65864 -0.00000 0.00000
12 0.00000 0.65862 0.00000 0.00000 0.00000 0.00000
This test stalls part way through when run with parallel execution. I am just running on a single machine using 2 or 4 cores in parallel, ARMCI_NETWORK=SOCKETS. The CPUs remain pinned at 100% but the job does not continue to make progress after about half of the CCSD energy calculations are complete. I first noticed this problem after the introduction of commit 25937 to tce_energy.F. If I revert that one file to the prior revision but otherwise leave the 6.5 beta code as-is, the li2h2_tce_ccsd test completes normally in parallel execution. Strangely none of the other TCE tests I have run stall in parallel. I don't know what is special about the interaction between this job and commit 25937.
|