parallel computing

zhangenze

Hi,

Does md.cluster.np means the number of the processor?

Can ISSM do parallel computing?

I change the value of np but it seems that nothing is changed. Do I need to have Matlab parallel toolbox to do parallel computing?

Best Regards,

Enze

mathieumorlighem

Hi Enze,

parallel computing is not supported yet for windows, but if you are using ubuntu or macOS, yes md.cluster.np is the number of processors used when you call md=solve(md,...).

Make sure to not request more processors than your machine has. Also, you will not notice any difference if your model is "small" (it could actually get worse). Generally speaking, it is not worth requesting more processors than 1/1000 elements. If you have 3,000 elements, you should try md.cluster.np=3;

I hope this helps

Best

Mathieu

zhangenze

Hi Mathieu,

I change the md.cluster.np from 1 to 2, and it generates errors.

Following are the error messages.

??? Error using ==> ./classes/Constraints/.././Constraints/SpcStatic.h

PenaltyDofAndValue error message: not implemented yet

If I change processor number to 16, it will repeat 16 times. I have almost 27000 elements. And I have 16 cores in total.

Could you help me to figure this out?

Best Regards,

Enze

mathieumorlighem

Hi Enze,

I don't know what you are trying to do but it is not supported. I quickly looked into the code and it seems like you are using Penalties and a flux-corrected method? Are you using moving boundaries with stabilization=4?? If this is the case this is under development and not supported.

Mathieu

zhangenze

Hi Mathieu,

I am using the flux-corrected method. Thanks for noticing. But if I only use one core, it still can work. So this method is not supported for just parallel computing, the flux-corrected method itself can work. Am I right?

Best Regards,

Enze

mathieumorlighem

To be honest, you should get the exact same error message with np=1, there is nothing special about parallel versus serial. We will probably remove the FCT method in the future, so you might want to switch to stabilization=1 which is more robust.

Best,

Mathieu

zhangenze

When I set np=1 and use the FCT method, it works. But it will take a very long time.

Best Regards,

Enze