|
[Sponsors] |
December 1, 2008, 18:31 |
Partitioning
|
#1 |
Guest
Posts: n/a
|
I recently read the "Advice on Using ANSYS CFX in Parallel" section of the CFX_11 Solver Guide and found a recommendation that the "percentage of overlap nodes to the total number of mesh nodes ideally be less than 10% for efficient partitioning". The guide also states that "values greater than 20% will impair performance".
Are these restrictions for each individual partition or for the sum of the partitions? I use Metis for the partitioning and I often get individual partitions with overlap percentages greater than 20% but the total for the sum of the partitions always remains less than 10%. I did some profiling and found that I got nearly linear speed-ups even though I had individual partitions with overlap percentages greater than 20% but there was one case (3.8 million elements, 1.4 million nodes) where I went from 2 to 4 processors and saw the code slow down. Any advice? Tristan |
|
December 1, 2008, 21:42 |
Re: Partitioning
|
#2 |
Guest
Posts: n/a
|
Hi,
I find the documentation a bit pessimistic here - I think you can use higher overlaps than they mention and for most simulations it's not too bad. Usually the parallel efficiency of the computer you are running on is much more significant, for instance on an Intel quad core you will probably get a speed up of 1.8 at dual processor but only 2.2 at quad. Any small parallel efficiency loss due to partitioning is irrelevant compared to this huge performance loss from the CPU architecture. Glenn Horrocks |
|
December 2, 2008, 12:28 |
Re: Partitioning
|
#3 |
Guest
Posts: n/a
|
Glenn,
So you believe the overlap limits listed in the documentation (<10% good, >20% bad) are supposed to apply to every partition and not just the sum of the partitions (even though they are pessimistic as you mentioned)? Thanks, Tristan |
|
December 2, 2008, 13:36 |
Re: Partitioning
|
#4 |
Guest
Posts: n/a
|
It really only applies to the entire mesh.
Think if it this way: a very small domain may not partition efficiently, resulting in >20% overlap. However, since it is a small domain it will not contribute significantly to the overall run time, nor will it contribute much to the overall overlap. An inefficiently partitioned larger domain will contribute more to the overall run time as well as the net overlap. If you have a lot of small domains, try using one of the "Coupled Partitioning" options. These options partition the entire mesh as though it were one domain, rather than dividing up each domain individually. This is a good approach if all your domains are fluid or porous domains, but not a good idea if you have solid domains, because solid domains only solve for the energy equation (plus solid AV's, if you have them), thus it is more efficient to distribute the solid domains evenly with the fluid ones. -CycLone |
|
Thread Tools | Search this Thread |
Display Modes | |
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
Partitioning problems in the parallel job | Neo | FLUENT | 5 | January 26, 2021 11:10 |
Weight in graph of mesh partitioning algo. | Joseph | Main CFD Forum | 9 | January 31, 2017 05:53 |
Mesh partitioning | krrishcfd | ANSYS Meshing & Geometry | 3 | May 16, 2010 22:33 |
Cicumferential Partitioning in CFX-11.0 | Khaled M. Essa | CFX | 2 | August 16, 2008 14:05 |
Using load vector for parallel partitioning | Kevala | FLUENT | 0 | May 1, 2001 13:11 |