Some questions about parallel simulation performace

Discussions about the Version 4 concerning the core model simulations
Lance
Posts: 12
Joined: Wed May 15, 2019 9:12 am

Some questions about parallel simulation performace

Postby Lance » Wed Sep 11, 2019 4:00 am

Dear ENVI-met team,

I have just got a high performance computer with 40-core CPU and 256 RAM and tried to use the science version of ENVI-met to test the parallel simulation performace. A test model was established with grids of 200 x 200 x 32 and the resolutions were set to 8m, 8m and 12m respectively. I recorded the CPU usage of different simulation periods which shows below.

1. Calculating 3D View factors (Parallel): CPU: 61%-37%
2. 3D initialization for all variables… (Initiating wind field…): CPU: 2%
3. 3D initialization for all variables… (Calculating new turbulence…): CPU: 40%
4. ENVI-met main loop: Writing initial files…: CPU: 7%-27%-20%
5. Start simulation: CPU: 20%-27% 10%-15% 61%-31%

During the whole simulation process, the maximum CPU usage comes to 61% and during the period 5, the CPU usage varies from time. It seems that the CPU does not take full use of performance in some of the time. I also ran two simulations at the same time and the maximum CPU usage still stays at 61%. So I have some questions about the performance of parallel simulation as follows.

Q1: Is there anything I can do to speed up the simulation since the CPU does not take full use of the performance?
Q2: Will the commercial version of ENVI-met run faster than the science version in parallel simulation?
Q3: Which method do you think will be more efficient for the large domain simulation, runnning one simulation or running two simulations at the same time?

Looking forward to your reply.

Regards,
Lance

Tim
Posts: 692
Joined: Mon May 15, 2017 3:46 pm

Re: Some questions about parallel simulation performace

Postby Tim » Wed Sep 11, 2019 7:11 am

Hi Lance,

For such a huge amount of CPU cores, you definitely need a larger model area. For model areas of 200x200x30, you probably find that a 8 core CPU with more GHz will be faster than this high performance PC, since splitting the calculations on the cores takes time as well.
You may try your tests with a 500x500x50 model area and check again. In general, I find that all simulations on every PC showed 100 % CPU usage at least for Calulating the 3D View factors or the radiation calculations. Might be only for a short time if that many cores are working on it but it should be that way when starting this calculation phases.
Furthermore, if you run two simulations at the same time and both are e.g. at calculating the View Factors, it has to be more than 61% usage! Something might be wrong in the software settings for your PC which does not allow ENVI-met to take full CPU usage? Or maybe there is a hardware setting which does not allow the PC getting too hot and thus sets a limit in CPU usage?

Q1: You cannot speed up the simulation by checking any additional speed factors.
Q2: No, they are the exact same version. Only the license type differs.
Q3: For that rather small area, where your PC does not take full CPU usage, I would recommend running at least two simulations in order to use as much of the CPU power as possible.

Best regards,
Tim

Lance
Posts: 12
Joined: Wed May 15, 2019 9:12 am

Re: Some questions about parallel simulation performace

Postby Lance » Thu Sep 12, 2019 2:42 am

Dear Tim,

Thanks for your advice. Btw, I have another doubt about the relationship between CPU clockwise and number CPU cores. Which factor is more efficient for the simulation? Or both of them are equally important? Since the high performance computer I use has 2.4GHz CPU clockwise speed with 40 cores, do you think the clockwise speed match with number of CPU cores?

Regards,
Lance

Tim
Posts: 692
Joined: Mon May 15, 2017 3:46 pm

Re: Some questions about parallel simulation performace

Postby Tim » Thu Sep 12, 2019 7:50 am

Hi Lance,

The relationship is not that clear but we found that e.g. 8 CPU cores with 3.8 GHz can be faster than 40 CPU cores with 2.4GHz depending on the model area size and shape. The effort of splitting calculations onto a large amount of CPU cores makes sense for very large model areas. For smaller model areas such as 200x200x30 it seems to be better to run them on a PC with less cores but a high CPU speed.

Best regards,
Tim

Lance
Posts: 12
Joined: Wed May 15, 2019 9:12 am

Re: Some questions about parallel simulation performace

Postby Lance » Fri Sep 13, 2019 2:39 am

Dear Tim,

Thanks you for your help.

Regards,
Lance


Return to “ENVI-met 4 (Main Simulation Model)”

Who is online

Users browsing this forum: No registered users and 1 guest