Parallel processing, processors weirdly utilized

wannab_phd · March 22, 2016, 10:00

Hello, I'm fairly new to OpenFOAM, but have decided today to run a small 2D case of flow around a cylinder, I failed, but that has another reason. The thing is, I ran it with multiple processors, I have an i5 4210H, dual core, but actually 4 cores, right, and I pulled that decomposeParDict, and specified the division of my domain to (2 2 1), z iz my empty direction. Weird thing happened, as I thought my processors would wind up to 99-100%, and the case would be solved in minutes. They instead acted like this:

http://s9.postimg.org/oykr1t4gf/parallel1.png
http://s22.postimg.org/9lgvo4ntd/parallel2.png

Iterations ran in terminal when all 4 processors were fully engaged, and then pause, nothing crashed, it wasn't a crash. The PC responded, and Terminal also responded, it just paused... Then continued to another peak, and paused again. Pauses were irregular.

Can someone explain that? Because I can't.

Thanks!

chriss85 · March 23, 2016, 05:34

Do they have a similar amount of cells and shared faces? It is also advised to run the renumberMesh utility for some speedup due to optimized caching. The latter point is probably not related to your issue though.

wannab_phd · March 23, 2016, 06:16

I don't know if they have a similar amount of cells and shared faces. How do I check that? Better yet, how do I make it similar? I will try the renumberMesh, thanks!

Antimony · March 23, 2016, 06:47

Hi,

You can take a look at the log file from decomposePar to see the distribution per processor.

To do a better load balance, you could use scotch as the method instead of simple/hierarchical. You don't need to define any coefficients as the code looks at the mesh and partitions it in the way it thinks is best.

Cheers,
Antimony

wannab_phd · March 23, 2016, 08:16

Will do! BTW does it (can it) matter if it's all ran from a live USB stick?

March 22, 2016, 10:00	Parallel processing, processors weirdly utilized	#1
wannab_phd New Member Join Date: Nov 2015 Posts: 6 Rep Power: 11	Hello, I'm fairly new to OpenFOAM, but have decided today to run a small 2D case of flow around a cylinder, I failed, but that has another reason. The thing is, I ran it with multiple processors, I have an i5 4210H, dual core, but actually 4 cores, right, and I pulled that decomposeParDict, and specified the division of my domain to (2 2 1), z iz my empty direction. Weird thing happened, as I thought my processors would wind up to 99-100%, and the case would be solved in minutes. They instead acted like this: http://s9.postimg.org/oykr1t4gf/parallel1.png http://s22.postimg.org/9lgvo4ntd/parallel2.png Iterations ran in terminal when all 4 processors were fully engaged, and then pause, nothing crashed, it wasn't a crash. The PC responded, and Terminal also responded, it just paused... Then continued to another peak, and paused again. Pauses were irregular. Can someone explain that? Because I can't. Thanks!

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Explicitly filtered LES	saeedi	Main CFD Forum	16	October 14, 2015 12:58
simpleFoam parallel	AndrewMortimer	OpenFOAM Running, Solving & CFD	12	August 7, 2015 19:45
problem for parallel processing	minaret	OpenFOAM Running, Solving & CFD	14	January 20, 2015 00:41
OF Parallel Processing with Core i7 - How to Handle Hyperthreading	dancfd	OpenFOAM Running, Solving & CFD	5	March 5, 2012 22:43
Parallel Computing on Multi-Core Processors	Upgrading Hardware	CFX	6	June 7, 2007 16:54

March 23, 2016, 05:34		#2
chriss85 Senior Member Join Date: Oct 2013 Posts: 397 Rep Power: 19	Do they have a similar amount of cells and shared faces? It is also advised to run the renumberMesh utility for some speedup due to optimized caching. The latter point is probably not related to your issue though.

March 23, 2016, 06:16		#3
wannab_phd New Member Join Date: Nov 2015 Posts: 6 Rep Power: 11	I don't know if they have a similar amount of cells and shared faces. How do I check that? Better yet, how do I make it similar? I will try the renumberMesh, thanks!

March 23, 2016, 06:47		#4
Antimony Senior Member Join Date: Aug 2013 Posts: 407 Rep Power: 16	Hi, You can take a look at the log file from decomposePar to see the distribution per processor. To do a better load balance, you could use scotch as the method instead of simple/hierarchical. You don't need to define any coefficients as the code looks at the mesh and partitions it in the way it thinks is best. Cheers, Antimony

March 23, 2016, 08:16		#5
wannab_phd New Member Join Date: Nov 2015 Posts: 6 Rep Power: 11	Will do! BTW does it (can it) matter if it's all ran from a live USB stick?