The number of processors shall I use

Anna Tian · October 9, 2012, 08:33

Hi,

I run my .def by parallel processors of the company. May I ask how to decide the number of processors shall I use to speed up my simulation?

Thanks!

monkey1 · October 9, 2012, 10:49

Hi Anna,

when using CFX the recommendations I got from the CFX support for the number of CPU's are:
1 CPU for approximately 250.000 Cells. So if your mesh consists of e.g. 1.000.000 Cells then you would use (at least) 4 CPU's.
This is what Ansys told me about optimum repartition and scaling. Although I usually use 8 CPU's for 1 Mio. Cells.

evcelica · October 9, 2012, 15:48

It also depends on your HPC licences, If you have HPC packs you are going to use the licence whether you run 2, 4 or 8 processors, so you might as well use 8. Two HPC pack licences would do 32 processors, so choosing 9 would be a pretty bad idea. Its different for standard HPC licences though, they are per core.

It also depends on your CPU and memory channels, 4 cores per processor seems to be where my computer tops out (Sandy Bridge E). Using more yields no gain at all and would be a waste of expensive HPC licences.

Anna Tian · August 1, 2013, 12:47

Quote:

Originally Posted by monkey1

Hi Anna,

when using CFX the recommendations I got from the CFX support for the number of CPU's are:
1 CPU for approximately 250.000 Cells. So if your mesh consists of e.g. 1.000.000 Cells then you would use (at least) 4 CPU's.
This is what Ansys told me about optimum repartition and scaling. Although I usually use 8 CPU's for 1 Mio. Cells.

Where did you find these recommendations? I didn't find them in the CFX tutorial. In order to acceleration the simulation ,shouldn't we use more processors than you proposed?

ghorrocks · August 1, 2013, 19:08

Usually the amount of processors you use is limited by the number of licenses you have, and how you are sharing them with other people using the system at the time. If you paid for the license then you might as well use them. It is pretty rare that adding a parallel license slows the simulation down (but is possible in extreme cases).

So I would not worry about the guidelines, and just use as many processors as you can practically use.

Anna Tian · August 3, 2013, 10:25

Quote:

Originally Posted by ghorrocks

Usually the amount of processors you use is limited by the number of licenses you have, and how you are sharing them with other people using the system at the time. If you paid for the license then you might as well use them. It is pretty rare that adding a parallel license slows the simulation down (but is possible in extreme cases).

So I would not worry about the guidelines, and just use as many processors as you can practically use.

I'm using the license of university. So the license number is not a contraint. But i have limited number of processors. I have quite a lot of jobs to run. So I need to set a reasonable number of processor to each job to not only accelerate the single simulation but also run as much job at the same time as possible. So I need a good way to relate the number of processors to be used to the grids number (just don't waste any processors).

ghorrocks · August 4, 2013, 08:07

If you have 5 big simulations to run and you have 5 machines to run it on, you will find running 5 separate serial simulations is faster and more reliable than running the simulations 5-way parallel one after the other. Of course this assumes the machines are big enough to run the simulation by themselves and lots of other issues.

But the point of this comment is that running in parallel results in speed and reliability losses. It produces results for a single solution more quickly, but if you have lots of simulations to run it might be faster and more reliable to run serial (or multi-processor with a low processor count).

monkey1 · August 5, 2013, 03:32

AnnaTian:
"Where did you find these recommendations? I didn't find them in the CFX tutorial. In order to acceleration the simulation ,shouldn't we use more processors than you proposed? "

These Informations were shared during ANSYS Training courses. In the Documentation you will not fid them, although there are some ppt Sheets available somewhere showing the scaling abilities of CFX.

October 9, 2012, 08:33	The number of processors shall I use	#1
Anna Tian Senior Member Meimei Wang Join Date: Jul 2012 Posts: 494 Rep Power: 16	Hi, I run my .def by parallel processors of the company. May I ask how to decide the number of processors shall I use to speed up my simulation? Thanks! __________________ Best regards, Meimei

August 1, 2013, 19:08		#5
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,850 Rep Power: 144	Usually the amount of processors you use is limited by the number of licenses you have, and how you are sharing them with other people using the system at the time. If you paid for the license then you might as well use them. It is pretty rare that adding a parallel license slows the simulation down (but is possible in extreme cases). So I would not worry about the guidelines, and just use as many processors as you can practically use. D.B likes this.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
[snappyHexMesh] Number of cells in mesh don't match with size of cellLevel	colinB	OpenFOAM Meshing & Mesh Conversion	14	December 12, 2018 09:07
How to let the mesh motion solver just solve a small region near a moving boundary?	zhajingjing	OpenFOAM Running, Solving & CFD	9	April 28, 2016 05:15
[snappyHexMesh] Layers:problem with curvature	giulio.topazio	OpenFOAM Meshing & Mesh Conversion	10	August 22, 2012 10:03
[snappyHexMesh] snappyHexMesh aborting	Tobi	OpenFOAM Meshing & Mesh Conversion	0	November 10, 2010 04:23
[Commercial meshers] Trimmed cell and embedded refinement mesh conversion issues	michele	OpenFOAM Meshing & Mesh Conversion	2	July 15, 2005 05:15

October 9, 2012, 10:49		#2
monkey1 Senior Member Join Date: Jul 2011 Location: Berlin, Germany Posts: 173 Rep Power: 15	Hi Anna, when using CFX the recommendations I got from the CFX support for the number of CPU's are: 1 CPU for approximately 250.000 Cells. So if your mesh consists of e.g. 1.000.000 Cells then you would use (at least) 4 CPU's. This is what Ansys told me about optimum repartition and scaling. Although I usually use 8 CPU's for 1 Mio. Cells.

October 9, 2012, 15:48		#3
evcelica Senior Member Erik Join Date: Feb 2011 Location: Earth (Land portion) Posts: 1,186 Rep Power: 23	It also depends on your HPC licences, If you have HPC packs you are going to use the licence whether you run 2, 4 or 8 processors, so you might as well use 8. Two HPC pack licences would do 32 processors, so choosing 9 would be a pretty bad idea. Its different for standard HPC licences though, they are per core. It also depends on your CPU and memory channels, 4 cores per processor seems to be where my computer tops out (Sandy Bridge E). Using more yields no gain at all and would be a waste of expensive HPC licences.

August 4, 2013, 08:07		#7
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,850 Rep Power: 144	If you have 5 big simulations to run and you have 5 machines to run it on, you will find running 5 separate serial simulations is faster and more reliable than running the simulations 5-way parallel one after the other. Of course this assumes the machines are big enough to run the simulation by themselves and lots of other issues. But the point of this comment is that running in parallel results in speed and reliability losses. It produces results for a single solution more quickly, but if you have lots of simulations to run it might be faster and more reliable to run serial (or multi-processor with a low processor count).

August 5, 2013, 03:32		#8
monkey1 Senior Member Join Date: Jul 2011 Location: Berlin, Germany Posts: 173 Rep Power: 15	AnnaTian: "Where did you find these recommendations? I didn't find them in the CFX tutorial. In order to acceleration the simulation ,shouldn't we use more processors than you proposed? " These Informations were shared during ANSYS Training courses. In the Documentation you will not fid them, although there are some ppt Sheets available somewhere showing the scaling abilities of CFX.