Nicolas,

I’ve never written code to run on super-computers, but it would not surprise me if it’s not extremely difficult to replicate earlier machine output on later equipment. And reading about GCMs it appears that is the equipment market they require.

I do think it would be interesting to find out what a high end, but still basic modern server could do in this arena, something like an 8-core xeon, or 16 core recent amd (sorry the name escapes me) with 256gb+ ram if the teams were to move away from antiquated programming methods. Be a pita to start from scratch though.

Models tend to react to stepwise jumps in the boundary conditions in interesting ways

Garry, have you considered Gerry Browning?

A small perturbation of the initial conditions for this equation can lead to instantaneous, unbounded growth and time dependent systems that exibit this type of behavior are called ill-posed systems. It is quite surprising how often simplifications that have been made in practice have led to this type of problem. Therefore, any simplification of the original continuum equations should be checked to ensure that the simplified system accurately approximates the continuum solution of interest and is properly posed

Link

This sounds like something you should try to correct. The only way a program with no user input can get different results in different runs is through some sort of pseudo random input. I’m sure that you do that to simulate noise, but you should be able to reproduce the identical pseudo random input if for no other reason than quality control. Otherwise how do you do regression testing?

Better _very_ late than never…

JF

Soon we will have an idea about the accuracy some of the parameters

Maybe not quite as soon as you think. See their draft policy on sharing data and results.

JF

OK

From a numerical analysis point of view, why does it happen?

I can’t speak for the UKMO folks, but we’ve had to release data with months set to all missing values not for numerical reasons (i.e., the model became unstable and produced unphysical values) but because the data wasn’t recoverable. As I noted, we’ve had data loss because of faulty magnetic tape and without a backup copy. Additionally, as hardware and software evolve, models fall by the wayside, as is the case with PCM. That model is no longer supported and cannot be used on currently-available hardware.

Also, as I said, machine changes, compiler changes, and OS changes can result in the loss of bit-for-bit reproducability. Given the way we run our climate models, rerunning an simulation can give different answers – not a different climate, but non-identical values over the length of the run.