partitioning advice

tkirkland · April 4, 2014, 18:07

I am working on modelling a screw compressor utilizing the one mesh per time step method. The meshes are generated outside of ansys by a specialized program. Each screw is a distinct domain and they share a GGI interface.
I am having no real issue running the simulation but would like to speed it along if possible. A run takes about 10 days.

My issue is.. when I partition it to run on multiple processors (I have 16 available) the %ovlp increases to 80% (@ 14 cores). I've tried almost every partitioning scheme out there with minimal reduction. The best solution I have come up with so far is to just run on fewer cores (22% on 2 cores, 60% on 6). If I delete this interface, overlap is <4% and it runs very quickly. So I know that is where my problem lies.

I have tried "part mmesh intersect option" but it didn't help.

I have about 3.5 million nodes right now but need to increase that a bit for other reasons.

Any ideas?

ghorrocks · April 6, 2014, 08:45

Does it matter? Probably not. I would do a short version of your simulation which runs in a few minutes (maybe a degree or so of rotation) and benchmark single processor versus multi processor up to 16. If you are getting good scaling up to 16 then who cares what the partitions look like. It you are getting poor scaling it will tell you how fast you can go effectively.

April 4, 2014, 18:07	partitioning advice	#1
tkirkland New Member tim kirkland Join Date: Mar 2014 Posts: 2 Rep Power: 0	I am working on modelling a screw compressor utilizing the one mesh per time step method. The meshes are generated outside of ansys by a specialized program. Each screw is a distinct domain and they share a GGI interface. I am having no real issue running the simulation but would like to speed it along if possible. A run takes about 10 days. My issue is.. when I partition it to run on multiple processors (I have 16 available) the %ovlp increases to 80% (@ 14 cores). I've tried almost every partitioning scheme out there with minimal reduction. The best solution I have come up with so far is to just run on fewer cores (22% on 2 cores, 60% on 6). If I delete this interface, overlap is <4% and it runs very quickly. So I know that is where my problem lies. I have tried "part mmesh intersect option" but it didn't help. I have about 3.5 million nodes right now but need to increase that a bit for other reasons. Any ideas?

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Partitioning problems in the parallel job	Neo	FLUENT	5	January 26, 2021 10:10
ERROR #001100279 has occurred in subroutine ErrAction.	smnaryal	CFX	11	December 20, 2017 16:32
OpenFOAM: looking for advice and compressible, nonisothermal tutorial preparation	bzz77	CFD Freelancers	1	February 2, 2016 12:11
Partitioning	Tristan	CFX	3	December 2, 2008 12:36
Advice on multi-phase flow modelling	Martin	Main CFD Forum	3	October 14, 2008 05:16

April 6, 2014, 08:45		#2
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,700 Rep Power: 143	Does it matter? Probably not. I would do a short version of your simulation which runs in a few minutes (maybe a degree or so of rotation) and benchmark single processor versus multi processor up to 16. If you are getting good scaling up to 16 then who cares what the partitions look like. It you are getting poor scaling it will tell you how fast you can go effectively.