What factors most affect computation speed - FPU performance, memory bandwidth ..??

Both, as well as memory latency being important, especially for AMD. It's really hard to say one is more important than the other, and it depends on how much you change one or the other.

An example from AMD64. I have AMD64 3400+ and 3800+ CPUs, each are running at 2.4 GHz with 512 KB L2 cache. Both run on Nvidia Nforce3 chipset motherboards. The latency of the memory modules on each PC are the same. The 3800+ has a dual channel memory controller and has significantly greater memory bandwidth because of it. The 3800+ does about 1.7 sec/TS while the 3400+ does about 1.9-1.95 sec/TS.

Why, are you thinking of upgrading your PC?

> Both, as well as memory latency being important, especially for AMD. It's
> really hard to say one is more important than the other, and it depends on how
> much you change one or the other.
>
> An example from AMD64. I have AMD64 3400+ and 3800+ CPUs, each are running at
> 2.4 GHz with 512 KB L2 cache. Both run on Nvidia Nforce3 chipset
> motherboards. The latency of the memory modules on each PC are the same. The
> 3800+ has a dual channel memory controller and has significantly greater
> memory bandwidth because of it. The 3800+ does about 1.7 sec/TS while the
> 3400+ does about 1.9-1.95 sec/TS.
>
> Why, are you thinking of upgrading your PC?
>

I am indeed thinking of upgrading my PC. The information on how the climate modelling uses resources is then obviously relevant, and more subtly also helps to optimise my current PC. So I have optimised memory bandwidth 'efficiency' to >90%, with only a very slight decrease in cycle time from 2.88 to 2.75 s/TS. One reads all sorts of benchmarks on CPUs and memory, but only rarely on complete systems for scientific applications, and certainly none for this one in particular. Also an issue here, with these very long computations, is (accumulated) computer errors; this may even be more important than pure speed. So, would ECC DRAM be an improvement for the project?

> helps to optimise my current PC. So I have optimised memory bandwidth
> 'efficiency' to >90%, with only a very slight decrease in cycle time from
> 2.88 to 2.75 s/TS.

Actually, that's not quite right. You've apparently massively improved your cycle time. If you take the latest trickle duration in seconds, and divide by 10802 timesteps (the number in a trickle), you get about 2.56 sec/TS. This is very good performance for a 2800+. The sec/TS on your trickle page is that accumulated over the entire model. So if you upgrade, or change something performancewise on your PC during a given model, it may show up gradually through continually decreasing sec/TS (as is happening with you now). If you changed it after the the 2nd trickle of the model, you would see significantly decreasing sec/TS over the next several trickles until it would even out about 2.55 to 2.6. If you changed it after the 30th trickle, there would instead be a slowly decreasing sec/TS for each trickle.

> Also an issue here, with these very long
> computations, is (accumulated) computer errors; this may even be more
> important than pure speed. So, would ECC DRAM be an improvement for the
> project?

It may, but I am unsure of that. Testing your system for a long period of time (say 12 to 24 hours) with memtest86+ and Prime95 should indicate whether there are any problems with your memory.

Related to the original question :

<a href="http://www.climateprediction.net/board/viewtopic.php?t=3148">http://www.climateprediction.net/board/viewtopic.php?t=3148</a>

<a href="http://www.climateprediction.net/board/viewtopic.php?p=29513">http://www.climateprediction.net/board/viewtopic.php?p=29513</a>