Dear all,

I am currently running 6 CPU's in paralel using the NET protocol. After 4 CPU's, I do not see any additional speed-up in time per iteration.

I switched to NMPI/Metis Partitioning, but this just made it run the same/slower.

The 6 CPU's are 3 Dual CPU machines, each with 4gb of ram, and on a gigabit network. We do not run out of ram/cpu on any of the machines.

Right now the cases are 3 million cells, steady segregated, 2nd order,Realizable K-Eps, and take about 25 seconds per itteration on the 6 CPU cluster...

Some advice please ~?!

I have noticed a similar problem with my simulations. My explanation (and I am no expert on the subject), a large amount of time is taken juggling data around between each computer node once you get above 4 CPU nodes.

Here are the stats that I found for my particular case:

I based efficiency on (CPU-time/Number of CPU's)/Wall Clock, also DT data transfer per iteration:

1 Node: 99.9873% 2 Nodes: 97.9519% (DT=0.025MB) 4 Nodes: 86.1218% (DT=0.063MB) 8 Nodes: 69.923% (DT=0.306MB) 16 Nodes: 40.512339% (DT=2.743MB) 32 Nodes: 16.906626% (DT=11.6MB)

I get around this by running a parallel version of Fluent 8 times, each taking 4 nodes.

Andrew

Yip, my problem is that I only have access to a single license, with 6 parallel licences.