[aspect-devel] Recommended number of cores for a model

[Rene Gassmoeller]

Hi Phil,

let me cc my answer to the mailing list, as this is a question I get
relatively frequently.


you can choose the number of processes for an ASPECT model quite
arbitrarily, but there are upper and lower limits on the number of
processors: The upper limit is given by the scaling of the (mostly
Stokes) solver, and for this model size of ~100 million DoFs my scaling
tests suggest to keep the number of Dofs per core at >=50,000 (i.e. at
max 2000 processes) to not waste compute time due to inefficient
solvers. The lower bound of number of processors is controlled by how
much RAM the nodes of your cluster have, and the wallclock time you are
willing to wait for your model. E.g. on a cluster with 2.5 GB RAM per
core (24 cores, 64 GB RAM) you should stay below 300,000 DoFs per core
(i.e. at least 333 processes for your model), otherwise the model might
crash, or run extremely slowly if your cluster allows swap space on the
hard disk. So for this example anything between 330 and 2000 processes
should be fine, but for different models and clusters this might be
different.


Best,

Rene


On 06/20/2017 11:26 AM, HERON, PHILIP J. wrote:
> Hi Rene, 
>
> I see in the paper you used 1535 processes for the model - is there a
> minimum processes you would recommend? Or is 1535 the minimum you used? 
>
> Cheers, 
>
> Phil
>
> Philip J. Heron
> Junior Research Fellow
> Dept. of Earth Sciences
> Durham University 
> web: http://philheron.com <http://philheron.com/>

-- 
Rene Gassmoeller
http://www.math.colostate.edu/~gassmoel/

-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.geodynamics.org/pipermail/aspect-devel/attachments/20170620/23d066b7/attachment.html>