Optmum number of nodes

Discussions centered around methodologies for high performance computing.

Optmum number of nodes

Postby chawas » Wed Oct 05, 2016 2:05 am

What will be the optimum number of nodes for running wrf.exe for two domains as shown by part of the namelist below.


&domains
time_step = 60,
time_step_fract_num = 0,
time_step_fract_den = 1,
max_dom = 2,
e_we = 508, 340,
e_sn = 463, 301,
e_vert = 30, 30,
p_top_requested = 5000,
num_metgrid_levels = 32,
num_metgrid_soil_levels = 4,
dx = 12000, 4000,
dy = 12000, 4000,
grid_id = 1, 2,
chawas
 
Posts: 6
Joined: Thu Sep 15, 2016 6:21 am

Re: Optmum number of nodes

Postby kwthomas » Thu Oct 06, 2016 1:53 pm

There are no rules for the number of nodes, nor will there every be. All supercomputers are different. They have different core counts per node. They have different amounts of memory per node.

Here are some general rules:

o If WRF dies and doesn't say anything useful on why it failed, it is likely that the job OOM'd
(Out Of Memory). In this case, you obviously need more nodes.

o Use all the cores per node, if possible. Idle cores mean slower performance. WRF wll
do all the right things if nproc_x/nproc_y don't divide evenly into your domain, though
that doesn't work for me. The CAPS ARPS postprocess software requires they divide
evenly.

o For best performance, nproc_x should be small. We're talking 4, 5, 6. 2 is too small. 3
will work, but may not be as efficient. This is based on testing on many different
supercomputers.
Kevin W. Thomas
Center for Analysis and Prediction of Storms
University of Oklahoma
kwthomas
 
Posts: 168
Joined: Thu Aug 07, 2008 6:53 pm


Return to High Performance Computing

Who is online

Users browsing this forum: No registered users and 1 guest

cron