same input data ,different output data!!!

[lxiaojuan8778@163_com]

:confused:I use cam3.1 to run two control runs on the same computer with same input datas ,but results are different slightly,especially to TS.There is a difference about 0.1K to 0.2K between the control runs.

Then I use IDEAL PHYS and found the difference disappeared.

Description of the Model said that :The new parameterizations compute the shortwave and longwave fluxes and heating rates for random overlap,maximum overlap, or an arbitrary combination of maximum and random overlap.

I want to know if error grew because of parameterizations ,and how can I remove this error. Should I change something in parameterizations or is ensemble experiments better ?

Thanks!!!!!!!!!

 

[eaton]

All versions of CAM are extensively tested to ensure reproducible results. If you are finding different results from multiple identical runs of cam3.1 then it's highly likely that there is a system problem.

The random overlap assumption referred to in the model description is implemented in a way that always gives identical results for identically configured runs. This is done by always seeding the random number generator the same way.

 

[lxiaojuan8778@163_com]

All versions of CAM are extensively tested to ensure reproducible results. If you are finding different results from multiple identical runs of cam3.1 then it's highly likely that there is a system problem.

The random overlap assumption referred to in the model description is implemented in a way that always gives identical results for identically configured runs. This is done by always seeding the random number generator the same way.

thanks for your reply!!
but if there is a system problem.Why IDEAL PHYS wipe off all the difference on the same server?
When I use PERGRO, I found the difference grows slower.

I use cam+som,not only cam ,and is it possible that parallel computing create this error??:confused::confused:Do you know what system problem may cause this problem??

thanks !!!

 

[eaton]

I used (misused?) the term "system" to include the compiler and mpi libraries.

If you're not getting reproducible results one of the first things to try is to reduce the compiler optimization.

If you're using smp mode then make sure that the answers are identical for different numbers of threads per task (i.e., the OMP_NUM_THREADS setting).

If you're using spmd mode then make sure the answers are independent of the number of MPI tasks used.

It's also possible that a combination of threading and compiler optimization is the problem. For example, it may be possible to run at a higher optimization without threading enabled than it is with it enabled.

If there is a compiler optimization problem, it is much more likely to show up with the full physics than with idealized physics. As you noticed the ideal physics has much slower error growth than the full physics. One reason is that there are no moist processes in the ideal physics.

 

[lxiaojuan8778@163_com]

I used (misused?) the term "system" to include the compiler and mpi libraries.

If you're not getting reproducible results one of the first things to try is to reduce the compiler optimization.

If you're using smp mode then make sure that the answers are identical for different numbers of threads per task (i.e., the OMP_NUM_THREADS setting).

If you're using spmd mode then make sure the answers are independent of the number of MPI tasks used.

It's also possible that a combination of threading and compiler optimization is the problem. For example, it may be possible to run at a higher optimization without threading enabled than it is with it enabled.

If there is a compiler optimization problem, it is much more likely to show up with the full physics than with idealized physics. As you noticed the ideal physics has much slower error growth than the full physics. One reason is that there are no moist processes in the ideal physics.

thanks a lot !!!!!!!!! I use -o0 and getting reproducible results !!!!