bug message reminder

When adressing a model crash or bug, please remember to include an entire model setup in your post that reproduces the crash or exposes the bug. Also add the XBlog.txt file. This is necessary information for people that are trying to help you. Including your model setup can be achieved by adding the zipped run directory (excluding output) as an attachment to the post.

Forum

RE: different results in MPI using 1 or 8 proc ???

CB
Cyprien Bosserelle, modified 7 Years ago.

different results in MPI using 1 or 8 proc ???

Capillary Posts: 0 Join Date: 11/22/11 Recent Posts
Hi,

I was playing around with a basic bathymetry and i found the wave height results were different whether I used 1 processor or 8 processors. Not much difference (~2% of the wave height )but still. I Narrowed down the problem to the advection scheme (with my grid ) where the central row of my grid had different yadvec. the error then propagates inshore. I narrow down the problem further to find that cgyv and the differences in fluxy had different values for the center row.

Can you reproduce this (see attached figure of the difference in H)?


I compiled XBeach with Ifort on ubuntu10. I simplified the problem to the maximum see param and bathy attached.

Cheers
Cyp
Bas Hoonhout, modified 7 Years ago.

RE: different results in MPI using 1 or 8 proc ???

Infra-gravity Posts: 362 Join Date: 5/20/11 Recent Posts
Thanks for the tip. This might be a bug. It has been a while that we compared MPI with non-MPI output. We will add your model to the skillbed and todo list, which is quite long...

Bas

Edit: moved to issue tracker: https://issuetracker.deltares.nl/browse/XBEACH-6
Robert McCall, modified 7 Years ago.

RE: different results in MPI using 1 or 8 proc ???

Wind Posts: 83 Join Date: 3/24/11 Recent Posts
Thank Cyprien for pointing this out, and for investigating where the problem might be. I went through the source code and it seems like there was a problem in communicating the dispersion relation across the internal MPI boundaries. With a code fix (implemented version 2780), the difference between 8-domain and 1-domain MPI models has been removed.

However, there will still be a small difference between non-MPI models (which use the stationary wave solver when instat='stat') and MPI models (which have to use the non-stationary wave solver when instat='stat'). We are trying to find ways of implementing the stationary wave solver in MPI, but this may still take a while.

Cheers,
Robert
LW
Luke Wright, modified 6 Days ago.

RE: different results in MPI using 1 or 8 proc ???

Capillary Posts: 1 Join Date: 1/14/20 Recent Posts
This is usually easy to visualize in I/O. However, I am also to define in the dissertation help UK blogs discuss version (MPI) of the same code and the difference is close to 0.3% in the error. The easy way of optimization, support, and added elements of value.