SUMMA parallelization

[AshlinAnn]

Hi, I am trying to set up SUMMA on Cray XC40. I am using a gfortran compiler. How do I enable parallelization for SUMMA? What changes should I bring in Makefile?

Thanks.

[wknoben]

Hi Ashlin,
Currently the most resource-efficient way to parallelize SUMMA runs is to use the built-in -g option. This option is used as follows:

startGRU = 1  # index of the GRU in the input .nc files; note that this refers to the index in the file (1..n) and NOT to the gruId variable
numGRU = 15   # number of GRUs to run, starting with startGRU
summa.exe -g startGRU numGRU -m filemanager.txt # runs SUMMA for 15 GRUs, starting with the first GRU in file

This will result in a split-domain output file, containing your requested outputs only for these 15 GRUs. The GRU indices for which the file has simulations are automatically included in the file's name.

You can access a brief description of what this does through summa.exe --help

[NicWayand]

Are there some more best practices for running SUMMA at scale (millions of points)?

[andywood]

Hi Nic, I typically use split domain runs with empirical load balancing of GRU ranges (ie with -g) across cores. It's clunky but works fine. We do a preamble run to determine the rough GRU ranges, then fix them in a joblist which can be run over many nodes. The new 'agents' stuff being built now may be a bit more streamlined and/or effective, but may also come with extra dependencies and overhead. It's probably worth benchmarking the different approaches for performance. I'd guess that w/o a formal MPI approach and more code optimization (ie parallel I/O & data objects) that we're reasonably close to what is achievable, efficiency-wise, versus a rudimentary domain-split.

…

On Mon, Jul 31, 2023 at 4:43 PM Nic Wayand ***@***.***> wrote: Are there some more best practices for running SUMMA at scale (millions of points)? — Reply to this email directly, view it on GitHub <#493 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABIKARIF2PCS3GSPHTP5MGTXTAYJNANCNFSM5F5HAYGQ> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[NicWayand]

I see there is a compile option with OpenMI, is that option working?

[NicWayand]

Thank you Andy. I will try the split domain approach for now.

[andywood]

apparently openMP works but doesn't scale all that well. I've never tried it versus the other more manual parallelization approaches, since I need to do those anyway with multi-node jobs

…

On Fri, Aug 11, 2023 at 3:53 PM Nic Wayand ***@***.***> wrote: I see there is a compile option with OpenMI, is that option working? — Reply to this email directly, view it on GitHub <#493 (reply in thread)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABIKARN4VFKL2QBFSPYEFOTXU2SUPANCNFSM5F5HAYGQ> . You are receiving this because you commented.Message ID: ***@***.***>