Low Cell Count Simulations

gigliagarf · September 13, 2015, 13:12

I'm running a interdymfoam case in openfoam with a few hundred thousand cells. Call it 300-500 thousand. Now I've read that there is a limit to the effectiveness of adding more cores, a few sources have stated anything less than 50,000 cells per core won't have any increased solve time. I've also read that there is a pretty good consensus that Intels chips are the way to go especially the I7's. So I'm currently solving this case with an older four core I7 but my solve time is around 24 hours. I'm looking to reduce the solve time, but I'm not too familiar with what to look for. From another source I've read that getting an 8 core machine can reduce my solve time by around 25 percent. But I'm hoping to reduce it further. What else should my computer have as far as hardware in order to more quickly solve this type of problem? I would appreciate your input. Of course if I've left out key details before you can fully answer the question just ask, I'll do my best to answer them. Thanks in advance, this website has been an invaluable resource so far.

flotus1 · September 13, 2015, 15:20

The limit for the cell count per core is rather a rule of thumb. For (some/many/a few) cases, parallelization efficiency decreases beyond this point.
Yet this does not mean that adding more cores can not increase simulation speed at all. Especially for small total cell counts run on a shared memory architecture you often see near linear scaling even for a much lower amount of cells per core.
And of course adding more cores is not the only way to increase simulation speed (CPU cache and frequency, memory frequency and bandwidth in general...)
Maybe before choosing new hardware you want to give more details about your current machine: exact CPU type, type and speed of RAM, number of DIMMs.
Are you running transient simulations where lots of data is written to disk during the simulation? Then also the hard disk type (HDD or SSD) could be important.

gigliagarf · September 13, 2015, 22:23

First of all thank you for you help! I've seen some of your replies before and you always give very informative answers. To answer your questions, it is a transient case. I'm running the I7-820 QM 1.73 GHz processor with 8 MB L3 cache. 4 GB of RAM over two slots of DDR3 PC3-10600 SDRAM running at 1333 MHz. All chugging on a HP Elitebook 8540w laptop. And I've changed to a solid state drive switching out the old HDD.

flotus1 · September 14, 2015, 08:11

Great! With such an outdated Laptop, virtually every modern 4-core workstation you buy will decrease simulation times by at least a factor of 2 (still a very conservative guess).

It depends on your budget which option is the best for you. If you went through the posts here you might already know that the Intel 2011-3 platform with an Intel I7-5820k and the fastest RAM possible (quad channel, dual rank) offers a pretty good price/performance ratio. Below that, Intel 1150 platform with an I7 processor and, you guessed it, the fastest ram you can afford (dual channel, dual rank) is the next best option.

gigliagarf · September 14, 2015, 19:52

Excellent! You've cleared up some of my misconceptions, thanks for the help.

September 13, 2015, 13:12	Low Cell Count Simulations	#1
gigliagarf New Member Join Date: Oct 2014 Location: Lafayette, LA Posts: 18 Rep Power: 12	I'm running a interdymfoam case in openfoam with a few hundred thousand cells. Call it 300-500 thousand. Now I've read that there is a limit to the effectiveness of adding more cores, a few sources have stated anything less than 50,000 cells per core won't have any increased solve time. I've also read that there is a pretty good consensus that Intels chips are the way to go especially the I7's. So I'm currently solving this case with an older four core I7 but my solve time is around 24 hours. I'm looking to reduce the solve time, but I'm not too familiar with what to look for. From another source I've read that getting an 8 core machine can reduce my solve time by around 25 percent. But I'm hoping to reduce it further. What else should my computer have as far as hardware in order to more quickly solve this type of problem? I would appreciate your input. Of course if I've left out key details before you can fully answer the question just ask, I'll do my best to answer them. Thanks in advance, this website has been an invaluable resource so far.

September 13, 2015, 15:20		#2
flotus1 Super Moderator Alex Join Date: Jun 2012 Location: Germany Posts: 3,427 Rep Power: 49	The limit for the cell count per core is rather a rule of thumb. For (some/many/a few) cases, parallelization efficiency decreases beyond this point. Yet this does not mean that adding more cores can not increase simulation speed at all. Especially for small total cell counts run on a shared memory architecture you often see near linear scaling even for a much lower amount of cells per core. And of course adding more cores is not the only way to increase simulation speed (CPU cache and frequency, memory frequency and bandwidth in general...) Maybe before choosing new hardware you want to give more details about your current machine: exact CPU type, type and speed of RAM, number of DIMMs. Are you running transient simulations where lots of data is written to disk during the simulation? Then also the hard disk type (HDD or SSD) could be important. gigliagarf likes this.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
How to determine the direction of cell face vectors on processor patches	sebastian_vogl	OpenFOAM Programming & Development	1	October 11, 2016 14:17
interFoam running blowing up	sandy13	OpenFOAM Running, Solving & CFD	2	May 5, 2015 08:16
Count breakup events per cell	a.weber	OpenFOAM Programming & Development	4	April 29, 2015 05:56
[snappyHexMesh] snappyhexmesh cell splitted ununiform..2D case	sharonyue	OpenFOAM Meshing & Mesh Conversion	1	April 23, 2013 00:34
How to determine the direction of cell face vectors on processor patches	sebastian_vogl	OpenFOAM Running, Solving & CFD	0	October 27, 2009 09:47

September 13, 2015, 22:23		#3
gigliagarf New Member Join Date: Oct 2014 Location: Lafayette, LA Posts: 18 Rep Power: 12	First of all thank you for you help! I've seen some of your replies before and you always give very informative answers. To answer your questions, it is a transient case. I'm running the I7-820 QM 1.73 GHz processor with 8 MB L3 cache. 4 GB of RAM over two slots of DDR3 PC3-10600 SDRAM running at 1333 MHz. All chugging on a HP Elitebook 8540w laptop. And I've changed to a solid state drive switching out the old HDD.

September 14, 2015, 08:11		#4
flotus1 Super Moderator Alex Join Date: Jun 2012 Location: Germany Posts: 3,427 Rep Power: 49	Great! With such an outdated Laptop, virtually every modern 4-core workstation you buy will decrease simulation times by at least a factor of 2 (still a very conservative guess). It depends on your budget which option is the best for you. If you went through the posts here you might already know that the Intel 2011-3 platform with an Intel I7-5820k and the fastest RAM possible (quad channel, dual rank) offers a pretty good price/performance ratio. Below that, Intel 1150 platform with an I7 processor and, you guessed it, the fastest ram you can afford (dual channel, dual rank) is the next best option.

September 14, 2015, 19:52		#5
gigliagarf New Member Join Date: Oct 2014 Location: Lafayette, LA Posts: 18 Rep Power: 12	Excellent! You've cleared up some of my misconceptions, thanks for the help.