hamlet at cs.umu.se
Wed Sep 13 00:46:33 PDT 2000
I've tried to run the xdlutime to get an estimate for how well our
cluster performs, but I can only test small matrices. I have 8 dual PIII
with 512 MB of memory so would at least be abel to run tests on
20000x20000, or am I wrong here?
dumburk [~/pfs]$ mpirun -np 8 -npn 1 xdlutime
Simple Timer for ScaLAPACK routine PDGESV
Number of processors used: 8
TIME N NB P Q LU Time Sol Time MFLOP/S Residual CHECK
---- ----- --- --- --- --------- --------- -------- -------- -------
Unable to perform LU-solve: need TOTMEM of at least 325494408
Bad MEMORY parameters: going to next test case.
On Tue, Sep 12, 2000 at 10:56:38PM -0400, Patrick GEOFFRAY wrote:
> Putchong Uthayopas wrote:
> > Below are our results. The machines is 8 Athlon 550MHz and one 1Ghz Athlon.
> > 512 Mb memory each. and Myrinet.
> > WALL 6000 9 1 9 116.80 0.41 1229.04 0.001097 PASSED
> > WALL 6000 9 9 1 200.79 3.65 704.66 0.001926 PASSED
> it seems that you are far from what you can get from your cluster
> * First, 9 nodes is not a good idea, the (1*9) or (9*1) grids are
> bad. you should use only 8 nodes and try (2*4) and (4*2) grids,
> much better.
> * Then 6000 is not a big matrix. If you have 512 MB per node, you
> should go up to 20000. It will run longuer but you will be closest
> to the peak.
> * Which Blas do you use ? Atlas gives a DGEMM peak at 600 MFlops
> per Athlon, so you can hope about 4 GFlops on your cluster.
> * Which MPI do you use ? if it's MPICH-GM, do you use the
> -gm-can-register-memory flag ?
> I strongly advice you to use the new High Performance Linpack
> (HPL) benchmark form Antoine Petitet
> (http://www.netlib.org/benchmark/index.html), it's faster than the
> old Linpack and much easier to install.
> Hope it can help.
> Patrick Geoffray
> Aerospatiale Matra - Sycomore
> Universite Lyon I - RESAM
> Beowulf mailing list
> Beowulf at beowulf.org
More information about the Beowulf