Difference between revisions of "WRF performance"

(Created page with "= Configuration = We tested the performance of WRF with the CORDEX domain with a nest over NSW, NARCliM project setup. It is possible to look at WRF performance with other do...")
 
Line 4: Line 4:
 
It is possible to look at WRF performance with other domains as long as those are used by a group of people and are used over a significant period of time if possible. Please contact [mailto:cws_help@nci.org.au cws_help@nci.org.au] to discuss the possibility.
 
It is possible to look at WRF performance with other domains as long as those are used by a group of people and are used over a significant period of time if possible. Please contact [mailto:cws_help@nci.org.au cws_help@nci.org.au] to discuss the possibility.
  
The configuration tested can be found [on Github
+
The configuration tested can be found at /g/data/sx70/wrf-scaling.
 +
 
 +
= Results =
 +
Below is a chart showing the walltime (WT) and the SU cost per model day for different number of processors, threads and I/O quilting options.
 +
 
 +
The chart shows the data for Gadi and some for Raijin as a comparison.
 +
 
 +
In the legend, "T" indicates the number of threads (OpenMP).
 +
 
 +
One can see for this configuration, the best option is to run with 480 processors, 2 OpenMP threads and 1 I/O group with 2 tasks. This gives the best walltime at one of the lowest cost in SU. Unfortunately, it seems simulations with I/O quilting aren't very stable (this doesn't show in those short tests) so please choose the configuration you think is best for you.
 +
 
 +
[[File:WRF scaling v4.1.3.png|1000px|center]]

Revision as of 20:14, 17 December 2019

Configuration

We tested the performance of WRF with the CORDEX domain with a nest over NSW, NARCliM project setup.

It is possible to look at WRF performance with other domains as long as those are used by a group of people and are used over a significant period of time if possible. Please contact cws_help@nci.org.au to discuss the possibility.

The configuration tested can be found at /g/data/sx70/wrf-scaling.

Results

Below is a chart showing the walltime (WT) and the SU cost per model day for different number of processors, threads and I/O quilting options.

The chart shows the data for Gadi and some for Raijin as a comparison.

In the legend, "T" indicates the number of threads (OpenMP).

One can see for this configuration, the best option is to run with 480 processors, 2 OpenMP threads and 1 I/O group with 2 tasks. This gives the best walltime at one of the lowest cost in SU. Unfortunately, it seems simulations with I/O quilting aren't very stable (this doesn't show in those short tests) so please choose the configuration you think is best for you.

WRF scaling v4.1.3.png