Closed milfeld closed 7 years ago
This issue persists with our local KNL system, Intel 17.1. I think the issue here is MULT=0 in the input. This would account for the -1 array size parameter when positive integer values of MULT are expected. The job ran fine when I changed it to a doublet. GAMESS bombs out with a descriptive message when a negative MULT is chosen, so I will track down the .LT.0 and change it to .LE.0.
Kris, Should I request developer access so that I can download code that has been "patched"? Thanks for answering this so quickly! (Please excuse my ignorance, Is it incorrect to request a multiplicity of 0 for this method?)
You shouldn't be able to choose MULT=0 for any method in GAMESS. The next public release will include the bug fix. Feel free to send in a request for developer access to get the patch sooner, but this bug is limited to jobs which shouldn't be able to start in the first place =)
I will keep this issue open a bit longer until I have time to make the fix. Thanks again for bringing this to our attention.
This has been fixed in the GAMESS development branch
Hi,
ModelA_cis_321g.txt
Hi all,
I can reproduce the problem on these platforms:
Stampede: intel 2015.0.2 compiler/MKL GAMESS VERSION = 1 MAY 2013 (R1) Centos 6.8
Lonestar5: intel 2016.0.1 compiler/MKL GAMESS VERSION = 1 MAY 2013 (R1) SUSE Linux Enterprise Server 11 SP3
Stampede KNL: intel 2017.0.0 compiler/MKL GAMESS VERSION = 18 AUG 2016 (R1) Centos 7.1
"mpi" build:
Execution with the attached input gives an MKL parameter error.
The error shows a wrong argument for DGEMV (stampede/lonestar5) and DGEMM (KNL). An argument for an array size appears as "-1"; even though the code continues to run and shows reasonable values for other processes (or may abort).
One would expect that the problem to be ILP/LP, but I don't suspect so because this is the first time we have seen this problem (on TACC machines), and the routines are being executed successfully by some of the other processes, and many other types of runs have not shown this.
I checked the lked commands and the log file reports:
vvvvvvvvvvvvvvvvvvvvvv This is the compile information: The message passing libraries searched are /work/00770/milfeld/knl_gamess3/gamess/ddi/libddi.a -L/opt/intel/impi/2017.0.098/lib64 -lmpi -lmpigf -lmpigi -lrt -lpthread
Other libraries, including math libraries, to be searched are -Wl,--start-group /opt/intel/mkl/lib/intel64/libmkl_intel_ilp64.a /opt/intel/mkl/lib/intel64/libmkl_sequential.a /opt/intel/mkl/lib/intel64/libmkl_core.a -Wl,--end-group ^^^^^^^^^^^^^^^^^^^^^^^^
The error messages are reported (with details from MKL_DEBUG=1, shown below as MKL_VERBOSE) as: Intel MKL ERROR: Parameter 2 was incorrect on entry to DGEMV . MKL_VERBOSE DGEMV(T,-1,26,0x2b4e260,0x2aabb0db8c18,457,0x2aabb0dcf120,1,0x2b4e268,0x2aabaa800b28,1) 63ns CNR:OFF Dyn:1 FastMM:1 TID:0 NThr:1
Could you possible determine if this a bug, or an incorrect build? Input attached:
Thanks. Best, Kent Milfeld Texas Advanced Computing Center (TACC)
login-knl1.stampede(13)$ cat info
!/bin/csh
compilation configuration for GAMESS generated on c562-014.stampede.tacc.utexas.edu generated at Sun Feb 19 10:05:30 CST 2017
setenv GMS_PATH ... setenv GMS_BUILD_DIR ...
setenv GMS_TARGET linux64 # machine type setenv GMS_FORTRAN ifort # Fortran Compiler setup setenv GMS_IFORT_VERNO 17 setenv GMS_MATHLIB mkl # mathematical library setup
setenv GMS_MATHLIB_PATH /opt/intel/mkl/lib/intel64 setenv GMS_MKL_VERNO 12 setenv GMS_DDI_COMM mpi # parallel message passing model setup setenv GMS_MPI_LIB impi setenv GMS_MPI_PATH /opt/intel/impi/2017.0.098 setenv GMS_LIBCCHEM false # LIBCCHEM CPU/GPU code interface setenv GMS_PHI knl # Intel Xeon Phi build: true/false setenv GMS_SHMTYPE sysv # Shared memory type: sysv/posix setenv GMS_OPENMP false # OpenMP support: true/false
Where for KNL (and the others): case 12: set MATHLIBS=" -Wl,--start-group ${MKLROOT}/lib/mic/libmkl_intel_ilp64.a ${MKLROOT}/lib/mic/libmkl_sequential.a ${MKLROOT}/lib/mic/libmkl_core.a -Wl,--end-group"