next up previous
Next: Reproducibility Up: PARALLEL MODEL VALIDATION Previous: PARALLEL MODEL VALIDATION

Error Growth

If the code is run on a different computer with even slight differences in machine arithmetic, the resulting output from the model will be different. The model is sensitive to initial conditions, and the particular path or trajectory taken by the model state represents the natural variability of the climate system. However, the rate of departure of the model state when started from slightly perturbed initial conditions is well known. A study by Rosinski and Williamson [29] indicates the expected departure in the temperature field, for example. A key component of the validation of the model implementation is the comparison of model output with CRAY model output. The CRAY model has been extensively compared with observational data and the characteristics (and shortcomings) of the model climate are documented in [20].

The following graphs give a comparison of diagnostic output from a CRAY YMP 1 day run and parallel runs on an 8x8 mesh on the Intel Paragon and an IBM SP2. The same T42 simulation was performed on each machine. The time step size is 20 minutes for a total of 72 timesteps per day. Figure 2 shows the root mean square of the temperature for the CRAY version of CCM2 and PCCM2 on the Paragon and SP2. The curves are indistinguishable graphically indicating several digits of agreement. Figure 3 shows the difference between the results for the parallel versions on the Paragon and the SP2.

A closer look at the other summary output from the model indicates the level of agreement of the implementations on different platforms. All these indicate very close agreement and are a substantial component of the validation of the implementations. Figure 4 shows the difference of the root mean square of the vorticity between the CRAY results and either the Paragon or SP2 results. Figure 5 shows the differences in the root mean square divergence and Figure 6 shows differences for the global moisture integral.

A comparison of history tapes at the end of a 1 day run was performed using the CCM post processor to analyze all the fields. The same level of agreement was found.



next up previous
Next: Reproducibility Up: PARALLEL MODEL VALIDATION Previous: PARALLEL MODEL VALIDATION



John B. Drake
Wed May 15 09:51:22 EDT 1996