Timing Information for the

NCAR CSM Ocean Model




Ocean modellers considering the NCOM will be interested in its timing requirements and cost comparisons with other global ocean-circulation models. This document provides a comparison of the timings for various configuations of three such models presently used in NCAR research:

The most straightforward, basic measure of model timing comparison is the number of CPU seconds per timestep per gridpoint required by each model. The table below summarizes this information for a small subset of NCOM, POCM, and POP experiments.

From early timing studies, we know that the additional physical and numerical features added to the NCOM cause it to run approximately a factor of 1.9 more slowly than the GFDL MOM 1.1. From data in the table, we can directly compare the timing costs of two models based upon MOM 1.1: the x2 NCOM and the 2x2 POCM. The ratio of these timing costs, 1.21e-5 : 0.70e-5 , is approximately 1.7, and is comparable to the NCOM/MOM 1.1 timing ratio.

Note that the POCM timing information was obtained from a single- processor run. For a more accurate comparison between POCM and NCOM, the 5-10% improvement in timing for the NCOM on a single processor should be included.




Table 1. Model Timing Comparisons

Model Seconds/ Timestep/ Gridpoint Seconds/ Timestep Number of Processors Computer Hardware Model Physics Momentum Timestep (seconds) Acceleration Factors 1
x3(NCOM) 0.96e-5 1.81 8 YMP 1 3504.0 [10,100]
x3(NCOM) 2.13e-5 4.02 16 J90 2 1 3504.0 [10,100]
x3(NCOM) 2.55e-5 4.81 16 J90 2 2 3504.0 [10,100]
x2(NCOM) 1.21e-5 9.21 8 YMP 3 2400.0 [10,100]
x2(NCOM) 1.24e-5 9.42 8 YMP 4 1800.0 [ 1,1 ]
2x2(POCM) 0.70e-5 2.17 1 YMP 5 2160.0 [10,100]
23M(POP) 7.23e-5 256.0 64 T3D 3 6 7200.0 [ 1,1 ]

1 In synchronous integrations, the tracer and barotropic stream function timesteps are equal to the momentum timestep value. For asynchronous (accelerated) integrations, the numbers in square brackets indicate the accelerated integration timestep factors. For example, "3504.0 [10,100]" should be interpreted as follows:

momentum timestep = 3504* 1 = 3504 seconds
upper-ocean tracer timestep = 3504* 10 = 35040 seconds
abyssal-ocean tracer timestep = 3504* 100 = 350400 seconds
2 The computational speed of a Cray-J90 is approximately one-half that of a Cray-YMP.

3 The computational speed of a Cray-T3D is approximately one-tenth that of a Cray-YMP.




Table 2. Model Resolution

Model Number of Zonal Gridpoints Number of Meridional Gridpoints Number of Vertical Gridpoints Total Number of Gridpoints Number of Islands
x3 (NCOM) 102 74 25 188700 8
x2 (NCOM) 152 111 45 759240 8
2x2 (POCM) 182 85 20 309400 4 (big)
23M (POP) 384 288 32 3538944 n/a




Table 3. Model Physics and Numerics

Model Reference Number from Table 1. Restoring Boundary Conditions Bulk Forcing Isopycnal Mixing K-Profile Parameter- ization (KPP) Upwinding Coupled with Flux- Coupler Free Surface
x3 (NCOM) 1 yes no yes no no no no
x3 (NCOM) 2 no yes yes yes no no no
x2(NCOM) 3 no yes yes yes yes no no
x2 (NCOM) 4 no yes yes yes yes yes no
2x2(POCM) 5 yes no no no no no no
23M (POP) 6 yes no no no no no yes




Factors Which Can Affect Model Timing and Timestep Size




CPU Hours per Year of Ocean-only Synchronous Integration on YMP

Note: the number of cpu hours for an accelerated integration is obtained by dividing the number of CPU hours listed above by the surface tracer acceleration factor. The model calendar follows the surface tracer time step.




How to Compare x2(NCOM) and 2x2(POCM) CPU Hours

x2(NCOM)/2x2(POCM) grid factor = 2.45
2x2(POCM)/x2(NCOM) timestep factor = 0.90
x2(NCOM)/2x2(POCM) physics factor = 1.90

grid factor * timestep factor * physics factor = 4.2

Therefore, the scaled x2 NCOM cpu time is 33.62 / 4.2= 8 hours.

Return to the top of this document




Questions?

This web page is maintained by Nancy Norton njn01@ncar.ucar.edu

Last modified on 11 March 1997