Problem when link the libgoto and acml using PathScale compiler
Posted: Thu Aug 11, 2005 2:51 am
Thanks Mr. Janne Blomqvist for his help, now I have successfully compiled the VASP code using PathScale compiler. But now I meet another ploblem. I can not link the libgoto into the final execute program. The acml or the combination of atlas and acml are both ok, but they show lower effeciency than the combination of libgoto and acml.
Followings are some benchmarks.
The system is BaTiO3，PREC=high, K-mesh is 15x15x15. The test machine is sun fire v20z with dual Opetron 246 cpus, 4 GB memory. Two mpi processes are executed on the same machine.
ifort 9.0+libgoto 0.99-3+acml 2.6+amd scalapack 2.5:
LOOP+: VPU time 439.15: CPU time 515.32
pgi 6.0+libgoto 0.99-3+acml 2.6+amd scalapack 2.5:
LOOP+: VPU time 469.39: CPU time 537.79
PathScale EKOPath 2.2+acml 2.6:
LOOP+: VPU time 545.97: CPU time 558.67
PathScale EKOPath 2.2+atlas 3.6.0+acml 2.6:
LOOP+: VPU time 569.44: CPU time 585.11
when I try to link in the libgoto, the compile does not any error. But when I execute the vasp problem, the stdout shows:
running on 2 nodes
distr: one band on 1 nodes, 2 groups
vasp.4.6.28 25Jul05 complex
POSCAR found : 3 types and 5 ions
WARNING: for PREC=h ENMAX is automatically increase by 25 %
this was not the case for versions prior to vasp.4.4
WARNING: for PREC=h ENMAX is automatically increase by 25 %
this was not the case for versions prior to vasp.4.4
WARNING: for PREC=h ENMAX is automatically increase by 25 %
this was not the case for versions prior to vasp.4.4
LDA part: xc-table for Pade appr. of Perdew
POSCAR, INCAR and KPOINTS ok, starting setup
FFT: planning ... 1
reading WAVECAR
WARNING: random wavefunctions but no delay for mixing, default for NELMDL
p0_31794: p4_error: interrupt SIGSEGV: 11
Killed by signal 2.
Obviously, the error occurs when vasp goes to the iteration part. So I believe the error is come from the incompatibility between libgoto and pathscale. This error still exists when I set the optimizing parameter to OPT=-O1.
Any one has met the similar problem? Can you give me any advice？Thanks in advance.
Gang Wu
BTW: When the optimization parameters are set to "-O3 -OPT:Ofast -OPT:fast_math=on -LNO:fu=9:full_unroll_size=7000 -fno-math-errno", and the blas and lapack are both acml, the vasp code compiled by the pathscale compiler will also show the waring message like "WARNING in EDDRMM: call to ZHEGV failed, returncode = 4 2 7". But the combination of ATALS and acml avoids this warring.
Followings are some benchmarks.
The system is BaTiO3，PREC=high, K-mesh is 15x15x15. The test machine is sun fire v20z with dual Opetron 246 cpus, 4 GB memory. Two mpi processes are executed on the same machine.
ifort 9.0+libgoto 0.99-3+acml 2.6+amd scalapack 2.5:
LOOP+: VPU time 439.15: CPU time 515.32
pgi 6.0+libgoto 0.99-3+acml 2.6+amd scalapack 2.5:
LOOP+: VPU time 469.39: CPU time 537.79
PathScale EKOPath 2.2+acml 2.6:
LOOP+: VPU time 545.97: CPU time 558.67
PathScale EKOPath 2.2+atlas 3.6.0+acml 2.6:
LOOP+: VPU time 569.44: CPU time 585.11
when I try to link in the libgoto, the compile does not any error. But when I execute the vasp problem, the stdout shows:
running on 2 nodes
distr: one band on 1 nodes, 2 groups
vasp.4.6.28 25Jul05 complex
POSCAR found : 3 types and 5 ions
WARNING: for PREC=h ENMAX is automatically increase by 25 %
this was not the case for versions prior to vasp.4.4
WARNING: for PREC=h ENMAX is automatically increase by 25 %
this was not the case for versions prior to vasp.4.4
WARNING: for PREC=h ENMAX is automatically increase by 25 %
this was not the case for versions prior to vasp.4.4
LDA part: xc-table for Pade appr. of Perdew
POSCAR, INCAR and KPOINTS ok, starting setup
FFT: planning ... 1
reading WAVECAR
WARNING: random wavefunctions but no delay for mixing, default for NELMDL
p0_31794: p4_error: interrupt SIGSEGV: 11
Killed by signal 2.
Obviously, the error occurs when vasp goes to the iteration part. So I believe the error is come from the incompatibility between libgoto and pathscale. This error still exists when I set the optimizing parameter to OPT=-O1.
Any one has met the similar problem? Can you give me any advice？Thanks in advance.
Gang Wu
BTW: When the optimization parameters are set to "-O3 -OPT:Ofast -OPT:fast_math=on -LNO:fu=9:full_unroll_size=7000 -fno-math-errno", and the blas and lapack are both acml, the vasp code compiled by the pathscale compiler will also show the waring message like "WARNING in EDDRMM: call to ZHEGV failed, returncode = 4 2 7". But the combination of ATALS and acml avoids this warring.