[Wrf-users] Execution Time Difference of ARW core versus NMM core

[Eric Altshuler]

Hello Mac,

Typically, the NMM core is 2-2.5 times faster than ARW for the same resolution, domain size and comparable physics options. For NMM to be slower (and 3 times slower at that) is very strange. First, you say you are using a 30x30 grid for both cores. Because of different grid staggering used by the two cores, a NxN grid in ARW will be square, while a NxN grid in NMM will be about twice as big in the X dimension as it is in the Y dimension. To get a square NMM domain, you need to use a (N/2)xN grid (for domains of approximately the same size, the values of N may be different for NMM and ARW, and the domains will not match exactly because of different map projections).

Also, are you still using a 72 second time step and getting those 'bad SMSTAV' error messages? Those messages indicate the value of SMSTAV is 'NaN' which means 'not a number' -- its value is so extreme that it can't be represented as a numerical value. Usually, as soon as any calculation in a program results in a NaN, the program will abort (usually with a floating point exception). However, depending on the type of machine and/or compiler options, it is possible for a program to continue running even after floating point exceptions have occurred. If this is the case with your NMM run, it may partly explain why it is so slow. As I suggested earlier, try reducing your time step to 26 2/3 seconds for NMM. If it is running properly, it should be significantly faster than ARW.

Best regards, 

Eric L. Altshuler
Assistant Research Scientist
Center for Ocean-Land-Atmosphere Studies
4041 Powder Mill Road, Suite 302
Calverton, MD 20705-3106
USA

E-mail: ela at cola.iges.org
Phone: (301) 902-1257
Fax: (301) 595-9793

----- Original Message -----
From: "mmacleod" <mmacleod at scotiaweather.com>
To: wrf-users at ucar.edu
Sent: Thursday, November 13, 2008 10:47:03 AM GMT -05:00 US/Canada Eastern
Subject: [Wrf-users] Execution Time Difference of ARW core versus NMM core

Good morning Folks.

We are setting up and testing the WRF cores for our company.  We have 
compiled both the ARW core and the NMM core on a small Dell dual core 
machine while we await the "big production system".

It is our understanding that the NMM core is about three times faster 
than the ARW core for the same configuration.

To test this we use the NAM 12 km in hourly outputs as the "driving model".

We set a 12 km 30 X 30 grid with 28 vertical levels.   We run the model 
to 6 hours. 

The OS is Ubuntu 8.4

We use the date  >  RUNTIME.LOG  command before the start of the wrf.exe 
module and the date >> RUNTIME.LOG to get the start and stop times.

Here is the results of the 6 hour runs:


NNM - 12 KM - 30 X 30 grid 28 levels  - 6 Hours

Thu Nov  6 19:55:49 GMT 2008
Thu Nov  6 20:04:03 GMT 2008


6 hr Run Time = 8 min 15 seconds

Run Time per hour = 1 min 22 seconds


ARW - 12 KM - 30 X 30 grid 28 levels  - 6 Hours

Thu Nov  6 20:34:00 GMT 2008
Thu Nov  6 20:35:38 GMT 2008


6 hr Run Time = 1 min 38 seconds

Run Time per hour = 27 seconds


This indicates that the NMM is three times slower than the ARW. 

Is this true in general?  If not what might be the cause of of our result?

Thanks in advance for your help.

Mac  

-- 
M.A. (Mac) MacLeod
President and General Manager
Scotia Weather Services Inc
192 Wyse Road, Suite 8,
Dartmouth, N.S. B3A 1M9

Tele: 902-468-3866
Fax:  902-461-1768
E-mail: mmacleod at scotiaweather.com
Visit us: www.scotiaweather.com

_______________________________________________
Wrf-users mailing list
Wrf-users at ucar.edu
http://mailman.ucar.edu/mailman/listinfo/wrf-users