Does the number of processors affect the simulation speed?
Dear All,
I'm conducting EDEM simulations with the various number of processors, for instance, 4, 8, 12, 16, 32 and 40 processors (CPU cores) in Unix computer of my university.
In my simulation, there are 518 steel particles of 1 mm in diameter and they are getting attacked by sonotrode that vibrates with 20kHz.
Even though I simulated it changing the number of processors, the simulation speed does not prominently change depending on the number of processors.
To finish 600 seconds of simulation time, it will take more than one week...
Any suggestion to significantly increase the simulation speed while remaining nice quality of the data would be helpful.
Thanks,
Nakwon
Answers
-
Hi Nakwon
It seems like your timestep might be too small. Could you tell me what your timestep is? Generally, you can use a value of 20% of the Rayleigh timestep to be the simulation timestep. Also, please check your save interval as to how frequently you are saving data from the simulation - that could potentially slow down your simulation also.
The other thing could be your simulation grid. The recommended value is 2-3 Rmin, so please check if that is within the recommended range.
These changes should make your simulation go faster. A 1-week simulation for only 518 particles with a 1 mm diameter seems like a lot, and this can definitely be improved.
Hope this helps.
Thanks,
Jerrin Job1 -
Jerrin Job Sibychan said:
Hi Nakwon
It seems like your timestep might be too small. Could you tell me what your timestep is? Generally, you can use a value of 20% of the Rayleigh timestep to be the simulation timestep. Also, please check your save interval as to how frequently you are saving data from the simulation - that could potentially slow down your simulation also.
The other thing could be your simulation grid. The recommended value is 2-3 Rmin, so please check if that is within the recommended range.
These changes should make your simulation go faster. A 1-week simulation for only 518 particles with a 1 mm diameter seems like a lot, and this can definitely be improved.
Hope this helps.
Thanks,
Jerrin JobHi, Jerrin Job
1. A value of time step is 12% of the Rayleight timestep to guarantee the reliability of the result data when the vibration is 20kHz and the amplitude is 50micrometer. So, even though I increase the time step, it still take long simulation time that is about a couple of days.
2. In Linux command, how can I regulate the save interval? Is that a command flag "-w" (write out interval)?
3. The grid size is 1.5Rmin in my case, so could it be a big deal to cause a quite long simulation time? Does the bigger grid size than the particle's diameter cause any additional negative effect on the simulation or the result?
Thanks,
Nakwon
0 -
Plus,
What are the effects of increasing the number of processors?
Doesn't it lead to the faster simulation speed?
0 -
Nakwon Kim said:
Plus,
What are the effects of increasing the number of processors?
Doesn't it lead to the faster simulation speed?
Hi Nakwon
More number of processors should make your simulation faster - that is correct.
Thanks,
Jerrin Job0 -
Nakwon Kim said:
Hi, Jerrin Job
1. A value of time step is 12% of the Rayleight timestep to guarantee the reliability of the result data when the vibration is 20kHz and the amplitude is 50micrometer. So, even though I increase the time step, it still take long simulation time that is about a couple of days.
2. In Linux command, how can I regulate the save interval? Is that a command flag "-w" (write out interval)?
3. The grid size is 1.5Rmin in my case, so could it be a big deal to cause a quite long simulation time? Does the bigger grid size than the particle's diameter cause any additional negative effect on the simulation or the result?
Thanks,
Nakwon
Hi Nakwon
-w is the flag for write out interval.
The grid size does not have any effect on the accuracy of the simulation. It only has effects on the simulation run time. If you have a very big grid, that can make your simulation slower. On the other hand, if you have a very small grid size, that can also make your simulation slower. So using a grid size of 3 R min is ideal. You can also use the 'Estimate Cell Size' option in the Simulator, to estimate the right size for your simulation.
Hope this helps.
Thanks,
Jerrin Job0 -
Jerrin Job Sibychan said:
Hi Nakwon
More number of processors should make your simulation faster - that is correct.
Thanks,
Jerrin JobHi, Jerrin Job
In the simulation with Unix computer remotely connected to my PC, if I use 40 CPU cores, it leads to more than 10000 hours to finish 600s of simulation, however, if I use 4 CPU cores, it takes about 345 hours to finish the simulation.
I do not understand why the decrease in the number of CPU cores results in the increase in the simulation speed.
Thanks,
Nakwon
0 -
Nakwon Kim said:
Hi, Jerrin Job
In the simulation with Unix computer remotely connected to my PC, if I use 40 CPU cores, it leads to more than 10000 hours to finish 600s of simulation, however, if I use 4 CPU cores, it takes about 345 hours to finish the simulation.
I do not understand why the decrease in the number of CPU cores results in the increase in the simulation speed.
Thanks,
Nakwon
Hi Nakwon
Are all the 40 CPU cores on the same node of your HPC? EDEM is unable to run across nodes, so I just wanted to check that with you.
Also, what is the OS of the system where you are running EDEM? EDEM can run on only select Linux OS that are listed here: https://community.altair.com/community/en/what-are-the-system-requirements-for-running-edem?id=kb_article&sysparm_article=KB0117125
Hope this helps.
Thanks,
Jerrin Job0 -
Hi Nakwon,
I just wanted to add that, given that you have very few particles in your simulation, the number of CPU cores used won't have a high influence in the simulation time.
Best regards,
Renan
0