This code is 2D, explicit, finite-difference, Navier-Stokes solver
written in Fortran (some Fortran 90).
The memory size is 564.8MB and I've run it on my 8-node, 16-processor
Xserve G5 cluster with 1GB/node. The software is XLF 8.1 and LAM-MPI.
The nodes are connected via Gig-E. Someone else ran the code using the
latest Myrinet/MPICH and the performance is exactly the same,
indicating that the code is not communication limited. This is not to
say that there isn't a lot of communication, but rather the code is
written such that while the non-blocking communication is taking place,
lots of computations which do not depend on the data being moved
around.
Notes:
1. From 2 to 4, and 8 processors, super-linear speed-up is achieved.
However its still less than 100% of original, 87% and 89%.
2. At 16 processors the scale-up is 72.7%, indicating that the
communications is starting to dominate.
3. This is good news for me as this is about 1/6th the size I typically
run this code at. I have a 3D version which requires roughly 40GB.
# CPUs Time(s)
1 1140
2 701
4 327
8 161
16 98
Sean
_______________________________________________
Do not post admin requests to the list. They will be ignored.
Scitech mailing list (email@hidden)
Help/Unsubscribe/Update your Subscription:
http://lists.apple.com/mailman/options/scitech/email@hidden
