[CPMD-list] Performance on beowulf cluster
Davide Provasi
davide.provasi at mi.infn.it
Mon Jul 7 12:06:18 CEST 2003
I am trying to optimize the preformance of CPMD (version 3.7.1)
on a beowulf cluster. The code is compiled with IFC Compiler
(Version 7.1, Build 20030521Z),Libraries: ATLAS 3.4.1,
LAPACK 3.0+updates, optimized for Pentium 4
My makefile contains the following flags:
> FFLAGS = -xiMK -pc64 -mp1 -ip -unroll -cm -w90 -w95
> -I/home/libs/mpich-1.2.5/include -safe_cray_ptr
> LFLAGS = -xiMK -Vaxlib -L/home/libs/ATLAS/Linux_P4/lib -lcblas -lf77blas
> -latlas -L/home/libs/LAPACK/libs -llapackP4
> -L/home/libs/ATLAS/Linux_P4/lib -lcblas -lf77blas -latlas
> -L/home/libs/mpich-1.2.5/lib -lmpich -safe_cray_ptr
> CFLAGS =
> CPP = /lib/cpp -P -C -traditional
> CPPFLAGS = -D__Linux -D__PGI -DLAPACK -DFFT_DEFAULT -DLINUX_IFC -DPARALLEL
> -DMP_LIBRARY=__MPI
> CC = gcc -O3 -Wall
> FC = ifc -c -tpp7 -O3
> #-safe_cray_ptr
> LD = ifc -tpp7 -O3
We have noticed a large Elapsed time in the FFTCOM,
and were wondering whether we could improve this.
****************************************************************
* *
* TIMING *
* *
****************************************************************
SUBROUTINE CALLS CPU TIME ELAPSED TIME
FFT-G/S 19040 291.46 292.76
S_INVFFT 4736 253.78 253.85
RHOOFR 54 103.04 102.86
S_FWFFT 1280 69.84 69.65
RGS 54 64.93 64.94
VPSI 114 43.68 44.02
EICALC 55 41.43 41.39
FFTCOM 6512 32.04 1756.83
FWFFT 275 25.34 25.28
GCENER 55 22.26 22.27
OVLAP 19 21.67 21.72
RNLSM1 150 21.02 21.11
INVFFT 221 18.21 18.26
FNONLOC 114 12.58 12.59
ROTATE 18 11.37 11.37
PCGRAD 18 10.21 10.54
GRADEN 55 7.76 7.94
PHASE 496 6.77 6.94
OVLAP2 96 5.00 5.82
XCENER 55 4.80 4.74
FORCES 18 3.30 3.37
----------------------------------------------------------------
TOTAL TIME 1070.49 2798.24
----------------------------------------------------------------
thank you in advance,
Davide
-----------------------------
Davide Provasi
University of Milano,
Physics Department
via Celoria 16, 20133 Milano
Phone +39 02 5031 7 717
-----------------------------
More information about the CPMD-list
mailing list