Why cores number in CFD simulation is preferred to be integer multiple of 4 for CFX?

Pierre1 · March 11, 2015, 21:31

I noticed that people usually put 4, 8, 16or 32 cores to run a CFX simulation. The cores number is always an integermultiple of 4. This seems to be a convention. I seldom see people use 11 or 9cores to run a CFX simulation.

I’m wondering does this convention make anysense? Why?

ghorrocks · March 12, 2015, 04:50

You are right, it does not make much sense. A 5 computer cluster will simply run about 25% faster than a 4 computer cluster. So feel free to use odd numbers if you wish.

I think it is simply because benchmarks are often done using factors of 2, so you expect an approximate halving of computational time. There is not much point benchmarking parallel efficiency on a 31 way cluster compared to a 32 way cluster - you will get far more meaningful results comparing a 64 way cluster to a 32 way one.

singer1812 · March 12, 2015, 16:46

Most cpus now days are quad cores. Probably the reason behind 4's. They just grab the whole chip when the run.

Our HPC is in clusters of 16, and it so happens that we have found that when we run with other software on the same node, it bogs everything down. Too lazy to trouble shoot, we usually just grab multiples of whole nodes when doing things.

Also, if you factor in hpc packs that ansys licenses, you will get another reason certain numbers of CPUs might be used.

QCFD · March 23, 2015, 18:03

In my experience, it is a hardware dependent choice. if a cluster is comprised of 12 core chips, it is useful to grab a multiple of 12 so you do not use a partial chip. If you have a partial chip and someone else tries to use the other cores you can have issues. In some cases, the other cores of the partially used chip are not made available to other users so the just sit idle.

ghorrocks · March 23, 2015, 20:20

For that to be the case it assumes that having multiple cores on a CPU working on the same job is better than having the multiple cores working on different jobs. Providing the jobs are equivalent (similar memory, similar parallelisation) in my experience it makes no difference what form the job takes it will load the system down about the same amount.

If you have a few cores idle then the active cores will run better as other bottlenecks (eg the memory and hard drive) are lighter loaded so the active cores work better. On top of that, many current Intel CPUs run at higher clock speeds when only one core is active which exaggerates this effect.

It is important to balance up the processes in a parallel run. If you have 4 identical 4 core CPU machines, and you send 4 processes to 3 machines but only 2 processes to 1 machine then it will not work too well as the loads are not balanced. It will run much better with either all machines having 4 processes, or even 3 processes.

March 11, 2015, 21:31	Why cores number in CFD simulation is preferred to be integer multiple of 4 for CFX?	#1
Pierre1 Member Pierre Join Date: Mar 2015 Posts: 37 Rep Power: 11	I noticed that people usually put 4, 8, 16or 32 cores to run a CFX simulation. The cores number is always an integermultiple of 4. This seems to be a convention. I seldom see people use 11 or 9cores to run a CFX simulation. I’m wondering does this convention make anysense? Why?

March 23, 2015, 18:03	Core count	#4
QCFD New Member Daniel Wilde Join Date: Jan 2014 Posts: 21 Rep Power: 12	In my experience, it is a hardware dependent choice. if a cluster is comprised of 12 core chips, it is useful to grab a multiple of 12 so you do not use a partial chip. If you have a partial chip and someone else tries to use the other cores you can have issues. In some cases, the other cores of the partially used chip are not made available to other users so the just sit idle.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Decomposing meshes	Tobi	OpenFOAM Pre-Processing	22	February 24, 2023 10:23
CFD Salary	CFD	Main CFD Forum	17	January 3, 2017 18:09
Help compiling code (newbie)	thsb20	OpenFOAM Programming & Development	56	August 24, 2016 05:24
CFD Design...The CFD Future	John C. Chien	Main CFD Forum	20	November 20, 2015 00:40
Reactor simulation using CFD takes to the Internet	Mark Brennan	Main CFD Forum	0	April 16, 1999 06:57

March 12, 2015, 04:50		#2
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,852 Rep Power: 144	You are right, it does not make much sense. A 5 computer cluster will simply run about 25% faster than a 4 computer cluster. So feel free to use odd numbers if you wish. I think it is simply because benchmarks are often done using factors of 2, so you expect an approximate halving of computational time. There is not much point benchmarking parallel efficiency on a 31 way cluster compared to a 32 way cluster - you will get far more meaningful results comparing a 64 way cluster to a 32 way one.

March 12, 2015, 16:46		#3
singer1812 Senior Member Edmund Singer P.E. Join Date: Aug 2010 Location: Minneapolis, MN Posts: 511 Rep Power: 21	Most cpus now days are quad cores. Probably the reason behind 4's. They just grab the whole chip when the run. Our HPC is in clusters of 16, and it so happens that we have found that when we run with other software on the same node, it bogs everything down. Too lazy to trouble shoot, we usually just grab multiples of whole nodes when doing things. Also, if you factor in hpc packs that ansys licenses, you will get another reason certain numbers of CPUs might be used.

March 23, 2015, 20:20		#5
ghorrocks Super Moderator Glenn Horrocks Join Date: Mar 2009 Location: Sydney, Australia Posts: 17,852 Rep Power: 144	For that to be the case it assumes that having multiple cores on a CPU working on the same job is better than having the multiple cores working on different jobs. Providing the jobs are equivalent (similar memory, similar parallelisation) in my experience it makes no difference what form the job takes it will load the system down about the same amount. If you have a few cores idle then the active cores will run better as other bottlenecks (eg the memory and hard drive) are lighter loaded so the active cores work better. On top of that, many current Intel CPUs run at higher clock speeds when only one core is active which exaggerates this effect. It is important to balance up the processes in a parallel run. If you have 4 identical 4 core CPU machines, and you send 4 processes to 3 machines but only 2 processes to 1 machine then it will not work too well as the loads are not balanced. It will run much better with either all machines having 4 processes, or even 3 processes.