Discussion Closed This discussion was created more than 6 months ago and has been closed. To start a new discussion with a link back to this one, click here.
Bad scalability
Posted 2010/01/15 0:22 GMT-5 4 Replies
Please login with a confirmed email address before reporting spam
Hello
I just tried how well comsol scale on a multicore processor. And the result is not very good...
The processor is an Intel core i7-975. 4 physical cores or 8 hyperthreaded cores and there are 6 GB of memory. I did the test using Vista 64-bit and comsol 3.5 64 bit.
The model is using the Navier stokes with k-e turbulence model (chns). the convection and conduction(chcc) and Maxwell stefan diffusion and convection (chms). All from the chemical engineering module.
I use the Paradiso solver to solve the time dependent solution (segregated).
Degree of freedom = 142610
I started the software with the -np flag and changed the value from 1 to 8.
The result:
NP Speed up
1 1
2 1.4733
3 1.4702
4 1.6667
5 1.5986
6 1.5606
7 1.5203
8 1.5155
Does anyone know why the result is so bad? I don't expect to get a linear speed up, but using 4 cores the speed up should be a bit higher. I guess it's a good idea to try a different solver? Does anyone have a tips of which solver to use. I don't have the time to test all of them.
I just tried how well comsol scale on a multicore processor. And the result is not very good...
The processor is an Intel core i7-975. 4 physical cores or 8 hyperthreaded cores and there are 6 GB of memory. I did the test using Vista 64-bit and comsol 3.5 64 bit.
The model is using the Navier stokes with k-e turbulence model (chns). the convection and conduction(chcc) and Maxwell stefan diffusion and convection (chms). All from the chemical engineering module.
I use the Paradiso solver to solve the time dependent solution (segregated).
Degree of freedom = 142610
I started the software with the -np flag and changed the value from 1 to 8.
The result:
NP Speed up
1 1
2 1.4733
3 1.4702
4 1.6667
5 1.5986
6 1.5606
7 1.5203
8 1.5155
Does anyone know why the result is so bad? I don't expect to get a linear speed up, but using 4 cores the speed up should be a bit higher. I guess it's a good idea to try a different solver? Does anyone have a tips of which solver to use. I don't have the time to test all of them.
4 Replies Last Post 2011/02/18 6:07 GMT-5
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hi
have you checked how much time you spend in pure number crunching (i.e true multiprocessing) and in model set-up, or redefinition, while you only run with on processor ?
Not everything "just" uses efficiently all 8 cores. Th effect is really very model complexity (and solver not all are foreseen for multicore) dependet
and are you sure you have enough local "scratch" files and RAM per core
Hope it helps to identify your issue, there have been a couple of dsussions about this
Ivar
have you checked how much time you spend in pure number crunching (i.e true multiprocessing) and in model set-up, or redefinition, while you only run with on processor ?
Not everything "just" uses efficiently all 8 cores. Th effect is really very model complexity (and solver not all are foreseen for multicore) dependet
and are you sure you have enough local "scratch" files and RAM per core
Hope it helps to identify your issue, there have been a couple of dsussions about this
Ivar
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
The result:
NP Speed up
1 1
2 1.4733
3 1.4702
4 1.6667
5 1.5986
6 1.5606
7 1.5203
8 1.5155
I agree, that does not look too impressive. One reason can be that you have a very small 2D model, where the assembly and factorization do not represent a significant part of the solution time. Small models do not scale very well.
Niklas
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hi,
When you say "small model do not scale very well", how small do you mean? (refered to degrees of freedom or which parameter?). I am preparing a 3D model which I have to be aware to be perfectly scalable, but it has to be small enough to be solved, because even with small scale (sizes around 5 mm x 5 mm x 5 mm), computer runs out of memory.
Thank you.
When you say "small model do not scale very well", how small do you mean? (refered to degrees of freedom or which parameter?). I am preparing a 3D model which I have to be aware to be perfectly scalable, but it has to be small enough to be solved, because even with small scale (sizes around 5 mm x 5 mm x 5 mm), computer runs out of memory.
Thank you.
Please login with a confirmed email address before reporting spam
Posted:
1 decade ago
Hello Niklas Rom,
In COMSOL 3.5a, I've noticed that it does not at all solve problems with high mesh resolution, probably because it runs out of storage space for the factorization matrices. Is there any way to overcome this using multicore? I have a workstation with 8 physical cores (16 hyperthreaded) and its amazing that COMSOL still cannot handle a fine mesh. I use the 2-D twophase Laminar Phase Field module.
Thanks,
Harish
In COMSOL 3.5a, I've noticed that it does not at all solve problems with high mesh resolution, probably because it runs out of storage space for the factorization matrices. Is there any way to overcome this using multicore? I have a workstation with 8 physical cores (16 hyperthreaded) and its amazing that COMSOL still cannot handle a fine mesh. I use the 2-D twophase Laminar Phase Field module.
Thanks,
Harish