Performance analysis of CP2K 2.6 vs 2.7

[Abhishek Bagusetty]

Hi Developers,

I have recently compiled a dev versions of CP2K and have been making some 
performance analysis. The test case is a GEO_OPT and here is the system 
size : 
Test Case specifications : 
Executable version : POPT
Runtype : GEO_OPT
System Size : 281 atoms
Nodes & Cores : 8, 128
Note : Exactly same nodes have been used for the test cases and there is no 
over commitment of the resources.

*Computational Walltime (hr:min:s)*
CP2K 2.6 w/o LIBSMM           : 4:57:45
CP2K 2.7(dev) with LIBSMM  : 6:11:04
CP2K 2.7(dev) w/o LIBSMM   : 5:42:42

The obvious reason is the differences in the time it takes for a *single 
SCF step in a given GEO Optimization step.*

*(Avg) Time for a single SCF step in any GEO Optimization step*
CP2K 2.6 w/o LIBSMM           :  3.5 s
CP2K 2.7(dev) with LIBSMM  :  4.8 s
CP2K 2.7(dev) w/o LIBSMM   :  4.7 s

There has been a significant difference in the walltime for the testcase. 
Has anyone experienced this sort of behavior ?

ARCH file  is shown below : 

CC       = CC

FC       = mpif90

LD       = mpif90 -lstdc++

AR       = xiar -r

DFLAGS   = -D__INTEL \

           -D__FFTSG \

           -D__parallel \

           -D__BLACS \

           -D__SCALAPACK \

           -D__FFTW3 \

           -D__LIBINT \

           -D__LIBXC2 \

           -D__HAS_smm_dnn           


CPPFLAGS = -C $(DFLAGS) -P -traditional


FCFLAGS  = $(DFLAGS) \

           -axSSE4.2 \

           -O2 -xW \

           -heap-arrays 64 -funroll-loops \

           -fpp -free \

           -convert big_endian \

           -I$(MKL_ROOT)/include \

           -I$(FFTW_ROOT)/include


FCFLAGS2 = $(DFLAGS) \

            -axSSE4.2 \

            -O1 -xW -heap-arrays 64 \

            -fpp -free


LIBSMM   = PATH_TO/build_libsmm


LDFLAGS  = $(FCFLAGS) \

           -L$(MKL_ROOT)/lib/intel64 \

           -L$(LIBINT_ROOT)/lib \

           -L$(LIBXC_ROOT)/lib \

           -L$(FFTW_ROOT)/lib \

           -L$(LIBSMM)/lib


LIBS      = -lmkl_rt \

            -lmkl_blas95_lp64 \

            -lmkl_lapack95_lp64 \

            -lmkl_scalapack_lp64 \

            -lmkl_blacs_openmpi_lp64 \

            -lfftw3 \

            -liomp5 \

            $(LIBSMM)/lib/libsmm_dnn_linux.intel.a \

            -lderiv \

            -lint \

            -lr12 \

            -lxc


OBJECTS_ARCHITECTURE = machine_intel.o

Thanks,
Abhishek
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20150625/470dbf80/attachment.htm>