SMP or NUMA scheme for Desktop

Fugacity · February 8, 2012, 11:34

My work was nice enough to purchase a new Dell t7500 workstation for CFD [and CAD work] because my laptop is not up to the task. It's a Dell T7500 w/ dual xeon X5647s and 54 GB of RAM right now. Personally I tried to setup an AWS Image but Remote Desktop is against our corp computer policy and is blocked by the firewall(s).

My question is, the BIOS lets me specify a SMP or NUMA RAM scheme. It's running SMP right now and seems fine; would/could NUMA have any tangible performance gains?

[and Yes, I did Google this topic. However, I found no satisfactory answers]

rmh26 · February 9, 2012, 13:58

It really depends on what software you are using.

For instance, lets say you have some large amount of data you are working on and you have parralized the code using openmp. Under a SMP system the data will be laid out in order in the memory in one large chunk. That memory might be controlled by the memory controller of processor A, ie its local to A. In order for processor B to access it, it will have to send a message to A asking for the data which will be sent back through the QPI (or HT for AMD). These interconnects are very fast at this point so I don't think the performance loss will be so great but it will depend on how compute intensive the operations are. In a NUMA system the data will be laid out in a scattered form so that part of the data is local to A and part of the grid is local to B. I think the memory is assigned on a First Use basis. So you might have to initially declare the data in a parallel manner. So maybe a little bit more work but less data sent between the processors.

If your using a commerical package my gut feeling is that SMP is fine but you could always run some benchmarks and post them up here to let people know.

February 8, 2012, 11:34	SMP or NUMA scheme for Desktop	#1
Fugacity New Member Chris Join Date: Jan 2012 Location: St. Louis Posts: 13 Rep Power: 14	My work was nice enough to purchase a new Dell t7500 workstation for CFD [and CAD work] because my laptop is not up to the task. It's a Dell T7500 w/ dual xeon X5647s and 54 GB of RAM right now. Personally I tried to setup an AWS Image but Remote Desktop is against our corp computer policy and is blocked by the firewall(s). My question is, the BIOS lets me specify a SMP or NUMA RAM scheme. It's running SMP right now and seems fine; would/could NUMA have any tangible performance gains? [and Yes, I did Google this topic. However, I found no satisfactory answers]

Thread Tools	Search this Thread
Show Printable Version Email this Page	Search this Thread: Advanced Search
Display Modes
Linear Mode Switch to Hybrid Mode Switch to Threaded Mode

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
how to understand high resolution scheme and high order scheme	iilw1314	Main CFD Forum	7	April 12, 2022 12:29
2nd order upwind scheme (Fluent and CFX)	Far	FLUENT	0	May 22, 2011 01:50
Godunov scheme	Peter	Main CFD Forum	6	November 27, 2009 10:51
Definition of limiter function for central dirrerencing scheme	sebastian_vogl	OpenFOAM Running, Solving & CFD	0	January 5, 2009 11:08
extrapolation in MUSCL scheme	Chandra	Main CFD Forum	6	February 14, 2007 11:21

February 9, 2012, 13:58		#2
rmh26 Member Join Date: Jul 2011 Posts: 59 Rep Power: 14	It really depends on what software you are using. For instance, lets say you have some large amount of data you are working on and you have parralized the code using openmp. Under a SMP system the data will be laid out in order in the memory in one large chunk. That memory might be controlled by the memory controller of processor A, ie its local to A. In order for processor B to access it, it will have to send a message to A asking for the data which will be sent back through the QPI (or HT for AMD). These interconnects are very fast at this point so I don't think the performance loss will be so great but it will depend on how compute intensive the operations are. In a NUMA system the data will be laid out in a scattered form so that part of the data is local to A and part of the grid is local to B. I think the memory is assigned on a First Use basis. So you might have to initially declare the data in a parallel manner. So maybe a little bit more work but less data sent between the processors. If your using a commerical package my gut feeling is that SMP is fine but you could always run some benchmarks and post them up here to let people know.