Hi. I've done some small test of my parallel network with the new Fluent 6.3. I wanted to figure out how fast the jobs were done comparing the number of CPU's used.
So first I did a serial run from an existing case with a data file. Only 320.000 cells though. I then completed 100 iterations logging the time. After that I did it in parallel based on the original dat file so the cases were completely identical. Much to my surprise the iteration plot was completely different for the two cases. The convergence level in the parallel job had much more scattering, and I was quite surprised to see that. Has anyone experienced the same? I suspect it has to do with the partitioning, but can anyone comment on that?
Note the two plots (last 100 iterations) serial and parallel. <table style="width:auto;"><tr><td></td></tr><tr><td style="font-family:arial,sans-serif; font-size:66%; text-align:right">From fluent6.3</td></tr></table> <table style="width:auto;"><tr><td></td></tr><tr><td style="font-family:arial,sans-serif; font-size:66%; text-align:right">From fluent6.3</td></tr></table>

I'd be willing to bet it's where your partitions were taken. If you do a contour plot of the residuals you'll probably see the maximum occuring where the partition passes through an area of noticeable curvature. You could try controlling the partition (instead of allowing the auto partition) to keep this from happening.

Hope this helps, and good luck, Jason

You're absolutely right!

The partitioning divides my sphere into two parts, with the line crossing the boundary layer in the separation zone.

You may choose a different partition method to partition the grid.