Port Validation in HP (formerly Compaq) Tru64

[sis@atmos_yonsei_ac_kr]

I have used WACCM (the Whole Atmosphere Community Climate Model,
which is based on CAM) on IBM POWER4 p690 cluster.

Recently, I have tried to get running the WACCM on Tru64 workstation.
I had no difficulty in compilation and got running the model.
Results on Tru64 workstation did not depend on the number of processors used.

Then, for validation, I did some perturbation growth tests based on the
documents posted on the CGD homepage.
However, any modifications of optimization levels of compilation on Tru64
workstation gave RMST difference of more than 10^{-2} after two day
simulation.

I'd like to know whether the Tru64 (OSF1) is appropriate to run the
CAM or not, since this effort might be useless if the Tru64 has some
problems.
In the past, before CAM, Tru64 (OSF1) was the supported platform.
Why the HP workstation based on Tru64 is not supported any more...?
Did that workstation have some problems?

 

[eaton]

The WACCM model is not currently a supported product (at least not by CSEG), though it may be in the future.

I have never tried the perturbation growth tests in WACCM and would be amazed if the test worked. As mentioned in the instructions for running the tests, CAM contains the CPP macro PERGRO which is used to modify algorithms that don't behave well in the perturbation test. This work has not been carried out for WACCM. The only way to validate the WACCM model is to do a climate simulation.

We are not currently testing CAM on a Tru64 machine since we don't have access to one. I'd suggest that a successful port validation of CAM on your machine would be a good sign that WACCM has been ported successfully as well. But again, a full climate simulation is the final validation.

 

[sis@atmos_yonsei_ac_kr]

Thanks for answers.

The pertrubation growth test in WACCM seems to work somehow, although
I commented out one line [qm1(...) = dqv(...)] within CPP PERGRO in
vertical_diffusion.F90 in both the AIX and Tru64 machine.

I agree with you. The port should be validated using a full climate
simulation using WACCM. Nonetheless, I think, some criteria to ensure
the validity of the port is still needed in case of using the full climate
simulation.

For example, even though climatology (e.g., 20-yr mean) obtained from a
Tru64 machine is consistent with that obtained from a AIX machine,
temporal behavior of field variables might not be so consistent with each
other, which results in inconsistent phase or amplitude.
In this case, is this port valid?

Criteria of a valid port is somewhat unclear to me.

Essentially, does a valid port mean almost identical results that are
different only at a few trailing digits?
Or, does even a valid port admit differences of results?

 

[eaton]

> I agree with you. The port should be validated using a full climate
> simulation using WACCM. Nonetheless, I think, some criteria to ensure
> the validity of the port is still needed in case of using the full climate
> simulation.

I believe the criteria to validate that two 20 year simulations are samples
from the same climate is an open research question (and possibly not a well
posed one). In practice one focuses on comparing those aspects of the
climate simulations that are of particular interest.

The value of the perturbation growth test is that it provides a simple way
to compare two simulations that are expected to differ only due to the
growth of machine roundoff size differences. This validation is extremely
useful for catching obvious problems with the port, before beginning the
much more difficult and expensive process of running long simulations and
doing a full climate validation using appropriate scientific criteria.

 

[sis@atmos_yonsei_ac_kr]

Thank you for reply

Sorry, I have another question. This will be the last one.

As far as I know, NCAR CGD provides officially CAM ported in the three
system (IBM, SGI, and LINUX).

What is the CGD's criteria to confirm whether the model is running
correctly in those systems? Are time-averaged fields and the perturbation
growth tests used? Is the transience of variables also considered?

And, is the port validated by comparing results with those obtained from
IBM system?

 

[eaton]

I'm not aware of an official support policy for CAM. The following
comments are my opinions/observations only.

The IBM is our production platform, and the model is tuned to produce an
acceptable climate there. The control simulations that are provided via the CAM
homepage are the final results of that tuning process.

CAM passes perturbation growth tests on our SGI (Origin 3800) and on an x86
based Linux cluster (using the Lahey Fortran compiler) compared to results
from the IBM. But we have not done full climate validations on either of
these platforms.

CAM running in the CCSM has been validated on more machines. See the CCSM
web site for a list of the fully supported machines on which full climate
validations have been performed. I don't know the details of that
validation process, but it does not use perturbation growth testing.