|
[Sponsors] |
August 14, 2019, 14:21 |
Serial vs Parallel Processing
|
#1 |
Member
yanki cob
Join Date: Jan 2018
Posts: 30
Rep Power: 8 |
Hello,
I have a workstation with 256GB RAM, 2 core Xeon 2.4GHz processors and Nvidia Quadro K620 GPU. Is it possible to run the simulations faster using parallel processing, or using gpu computing with serial processing? Bear in mind I have no idea what the difference is between the two. My model does not contain any UDF's, symmetry or periodic boundary conditions (if it matters). Regards, Yanki |
|
August 14, 2019, 20:03 |
|
#2 |
Senior Member
Svetlana Tkachenko
Join Date: Oct 2013
Location: Australia, Sydney
Posts: 407
Rep Power: 14 |
I have not really tried but I expect GPU to be a lot faster. To be sure I would suggest to test it out with one cpu, or one cpu and one gpu, or two cpus, or two cpus and one gpu. These options are configurable in the ansys fluent launcher dialog.
|
|
August 16, 2019, 03:28 |
|
#3 |
Member
Join Date: Mar 2016
Posts: 73
Rep Power: 10 |
I recommend that you do a little research. Fluent is capable of using GPU for solving, which is way faster than CPU. But as far as I know that doesn't work for all solvers and some solvers are less stable or less accurate with a GPU.
And you should find out if you GPU is supported. That's easy if you set up GPGPU in your launch window and check in Fluent if you GPU is supported. |
|
August 16, 2019, 05:54 |
|
#4 | |
Member
yanki cob
Join Date: Jan 2018
Posts: 30
Rep Power: 8 |
Quote:
"ANSYS ICEPAK supports NVIDIA's CUDA-enabled Tesla and Quadro series workstation and server cards." But, I don't know if I should install something or somehow activate it for this purpose. When I run a simulation with GPGPU, my GPU's performance in the task manager is still zero, so it doesn't work. |
||
August 16, 2019, 07:01 |
|
#5 | |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,400
Rep Power: 47 |
Objection!
GPU computing is not that much faster, especially not with a Quadro K620. And there are a lot of caveats when it comes to GPU computing. To be clear: OP will most likely get less performance with GPU acceleration activated. Instead, just use parallel processing on CPU with the maximum number of licenses and physical cores you have available. Quote:
Most of the solvers fluent uses scale linearly up to a certain amount of cores. Which means running parallel on 2 cores instead of 1 cuts wall time in half. Edit: the fact that the GPU stays idle even with GPU acceleration activated indicates that you are using one of the solvers that do not benefit from GPU acceleration at all. E.g. the Fluent default SIMPLE solver. Fluent knows this and ignores GPU acceleration for these solvers by default. Forcing it to use GPU anyway will lead to much lower performance. |
||
August 16, 2019, 07:40 |
|
#6 | |
Member
yanki cob
Join Date: Jan 2018
Posts: 30
Rep Power: 8 |
Quote:
That's what I exactly did. I thought I had 2 cores, but turns out there are 56 of them (don't know what they are called, you can find them on task manager). So, 56 cores for parallel processing, which makes everything faster. Regarding GPU, you are right. Also, I use SIMPLE, so it doesn't make sense to use GPU acceleration (as I've read). |
||
August 16, 2019, 08:03 |
|
#7 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,400
Rep Power: 47 |
There is a chance that you don't have 56 physical cores, but only 28 cores and 56 threads thanks to Hyperthreading/SMT. These virtual cores are useless for Fluent, often resulting in lower performance when using them. So knowing how many physical cores you have would be useful. For dedicated CAE workstations, SMT is usually turned off in the BIOS.
To find your CPU model, click the start-button then right-click on "computer". in the context menu click on item on the bottom, I think it's called settings or preferences. |
|
August 16, 2019, 10:06 |
|
#8 | |
Member
yanki cob
Join Date: Jan 2018
Posts: 30
Rep Power: 8 |
Quote:
Right now, I have 3 simulations running on this machine with 2 of them using 16 cores each and the last one with 4 cores, a total of 36 (70% CPU performance). Are you saying if I used a total of 28 cores, I would get a better fluent performance? edit: my CPU model is E5-2680 v4 and I have 2 of them hence 28 cores Regards, Yanki |
||
August 16, 2019, 10:23 |
|
#9 |
Super Moderator
Alex
Join Date: Jun 2012
Location: Germany
Posts: 3,400
Rep Power: 47 |
Yes, using more than 28 threads (i.e. the number of physical cores you have) in total will not give you any benefit in Fluent. And in many cases, using more than that will even slow down the simulation.
With SMT turned on, Windows task manager will report about 50% total CPU usage when running only 28 threads. Don't worry, that's just a flaw in the CPU utilisation readout. You are still using your hardware at its maximum potential. |
|
August 16, 2019, 10:26 |
|
#10 | |
Member
yanki cob
Join Date: Jan 2018
Posts: 30
Rep Power: 8 |
Quote:
|
||
January 14, 2021, 05:54 |
|
#11 |
New Member
Mahesh
Join Date: Mar 2020
Posts: 20
Rep Power: 6 |
Will the results vary much using Serial and Parallel processors? For me they are varying with a considerable extent. Please anyone explain the technicality behind if they are supposed to vary.
Thanks in advance. |
|
|
|
Similar Threads | ||||
Thread | Thread Starter | Forum | Replies | Last Post |
Case running in parallel gives error whereby running in serial works | Harak | OpenFOAM Running, Solving & CFD | 8 | October 17, 2015 11:12 |
Explicitly filtered LES | saeedi | Main CFD Forum | 16 | October 14, 2015 11:58 |
Parallel Processing by region | Swift | OpenFOAM Running, Solving & CFD | 0 | July 30, 2015 08:42 |
Parallel processing problem | newbie29 | OpenFOAM Running, Solving & CFD | 1 | June 22, 2012 04:23 |
Parallel processing and pitzDaily example | roncresswell | OpenFOAM Running, Solving & CFD | 1 | October 18, 2005 19:02 |