Gauss solution on the cluster:
system size -> n
RAM for matrix A -> r = 8*n^2/1e6
processor number -> z
RAM of matrix A on each node -> d = r/z
------------------------------------------------------------------
A on one node:
n = 3000 number of unknowns
z = 1 processor
r = 72 MBytes for whole matrix A
d = 72 MBytes of A on each node
mpirun -machinefile machi.dat -np 1 gauss3.exe
times for array with leading dimension 3000
factor solve total mflops unit ratio
6.760E+02 0.000E+00 6.760E+02 2.665E+01 7.504E-02 1.207E+04
end of test
------------------------------------------------------------------
B on three nodes
n = 3000 number of unknowns
z = 3 processors
r = 72 MBytes for whole matrix A
d = 24 MBytes of A on each node
mpirun -machinefile machi.dat -np 3 gauss3.exe
times for array with leading dimension 3000
factor solve total mflops unit ratio
3.200E+02 1.400E+01 3.340E+02 5.395E+01 3.707E-02 5.964E+03
end of test
------------------------------------------------------------------
C speed_up_{np = 3} = 5.395E+01 / 2.665E+01 = 2.0244
------------------------------------------------------------------
D corrida con el maximo tamanio del sistema posible en el
actual cluster:
n = 6500 number of unknowns
z = 3 processors
r = 338 MBytes for whole matrix A
d = 112.67 MBytes of A on each node
pghpf -Mautopar -Mmpi -o gauss3.exe gauss3.hpf
mpirun -machinefile machi.dat -np 3 gauss3.exe
times for array with leading dimension 6500
factor solve total mflops unit ratio
3.152E+03 3.800E+01 3.190E+03 5.742E+01 3.483E-02 5.696E+04
end of test
--
CimecUser - 01 Sep 2001
This topic: Main/Cimec > PGHPFPerformanceReport
Topic revision: 01 Sep 2001, CimecUser