Dual Nodes is Slower Than Single Node (Reposting)

Mrxlazuardin · May 22, 2010, 04:59

Hi,

I was posting this thread on Fluent forum. Now, I'm reposting to get different of view on my problem.

I don't understand why Fluent running with 32 processes on dual nodes is slower than Fluent running with 16 processes on single node? Each node has four Quad Core Opteron CPUs with eight 2GB RAMs and GbE NIC. I'm using Windows Server 2008 x64. I need suggestion how to optimize this dual nodes configuration.

Best regards,

PGodon · May 26, 2010, 11:25

I don't know a thing about FLUENT as I am writting my own codes... but there are two aspects that I can address.

First as you may know when making your code parallel there is part of the CPU time that goes into communication between the processors. As the number of processors increases, each processor spends less time doing its own task while the communication between processors increases. There is a treshold (for each problem it is different) for which increasing the number of processors will not improve the performance but rather it will slow it down, as the processors will lose time in communication. So here you might have reach that treshold for that particular problem. For very large grids in 2 and 3 D, many processors is good, but for smaller 2D grids sometimes 10 processors is best, above that the CPU needed increases...

Second, there are some hardwares, and I am not a pro - I am just repeating what I read somewhere, such that the duo and quadro are made of individual processors that are not as fast as some of the "single" (solo) processors. So it might also be due to a hardware specification, I would not know for your particular case, but I know it is for some computers.

And BTW, there are also now some intel 6-core chips available on the market and one can build a desktop with two of these to reach a dozen of processor in one little box.... better than these bulky clusters and cheaper... though of course for some the cluster might still be the solution.

May 22, 2010, 04:59	Dual Nodes is Slower Than Single Node (Reposting)	#1
Mrxlazuardin New Member Lazuardi Nasution Join Date: Apr 2010 Posts: 4 Rep Power: 16	Hi, I was posting this thread on Fluent forum. Now, I'm reposting to get different of view on my problem. I don't understand why Fluent running with 32 processes on dual nodes is slower than Fluent running with 16 processes on single node? Each node has four Quad Core Opteron CPUs with eight 2GB RAMs and GbE NIC. I'm using Windows Server 2008 x64. I need suggestion how to optimize this dual nodes configuration. Best regards,

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
Dual Nodes is Slower Than Single Node	Mrxlazuardin	FLUENT	0	May 21, 2010 02:48
Scale-Up Study in Parallel Processing with OpenFoam	sahm	OpenFOAM	10	April 26, 2010 18:37
Single vs Dual Processors	Sam Z	CFX	4	October 22, 2002 18:17
P4 1.5 or Dual P3 800EB on Gibabyte board	Danial	FLUENT	4	September 12, 2001 12:44
CFX4.3 -build analysis form	Chie Min	CFX	5	July 13, 2001 00:19

May 26, 2010, 11:25		#2
PGodon New Member Patrick Godon Join Date: Apr 2010 Posts: 19 Rep Power: 17	I don't know a thing about FLUENT as I am writting my own codes... but there are two aspects that I can address. First as you may know when making your code parallel there is part of the CPU time that goes into communication between the processors. As the number of processors increases, each processor spends less time doing its own task while the communication between processors increases. There is a treshold (for each problem it is different) for which increasing the number of processors will not improve the performance but rather it will slow it down, as the processors will lose time in communication. So here you might have reach that treshold for that particular problem. For very large grids in 2 and 3 D, many processors is good, but for smaller 2D grids sometimes 10 processors is best, above that the CPU needed increases... Second, there are some hardwares, and I am not a pro - I am just repeating what I read somewhere, such that the duo and quadro are made of individual processors that are not as fast as some of the "single" (solo) processors. So it might also be due to a hardware specification, I would not know for your particular case, but I know it is for some computers. And BTW, there are also now some intel 6-core chips available on the market and one can build a desktop with two of these to reach a dozen of processor in one little box.... better than these bulky clusters and cheaper... though of course for some the cluster might still be the solution.