[MITgcm-support] Problem of Parallel Processing

[Jody Klymak]

On Mar 29, 2014, at  7:00 AM, 连展 <lianzhan at fio.org.cn> wrote:

> Hi there
> 
> Does anyone have the experience of run the massively parallel computing with MITgcm, like hundreds of processes? 
> 
> I just read the manual, and I have two problems about this.
> 
> The first is: we can make one process is allocated to only 1 tile, or we can make several titles computed in one process. So what is the principle for the decomposition? I mean, how different it will be if we make "sNx=100, sNy=100, nSx=1, nSy=1, nPx=100, nPy=100" or "sNx=10, sNy=10, nSx=10, nSy=10, nPx=100, nPy=100"?

I think the latter will have a bunch of message passing and extra tile overlap data that is unnecessary and will slow you down.  I expect the only reason for subtiles (nSx,nSy>1) would be if you had a very complex domain and wanted to use the exch2 package to arrange your subtiles to not be over land. 

> 
> The second question is: If we decompose a domain with land and sea, there may be some tiles which is totally land. If we were running an ocean model, what would happen if the 100% land tile is allocated to a process? If the process has only one tile to be computed, would this process do nothing but waiting or this tile wil! l be excluded from the computation(like Fig 4.4 in the manual, the blanked tiles) and the process will get a new tile automatically?

I don't think this gets done by default, and I think can be done using the exch2 package.  I've not used exch2 for this purpose, but I hope to soon for some tidal channel modelling.

Cheers,   Jody



> Thanks very much!!!
> 
> 
> --
> 
> Lian
> _______________________________________________
> MITgcm-support mailing list
> MITgcm-support at mitgcm.org
> http://mitgcm.org/mailman/listinfo/mitgcm-support