My new WU http://climateapps2.oucs.ox.ac.uk/cpdnboinc/workunit.php?wuid=240083
has also been sent to four other computers (the first gave error). I'm curious about the reasons. Are you testing how the same model runs in similar systems (all the computers are AMD athlon XP with Windows XP)?

---

Professor Desty Nova
Researching Karma the Hard Way

> Are you testing how the same model runs in similar systems
> (all the computers are AMD athlon XP with Windows XP)?
>
There is discussion of this <a href="http://www.climateprediction.net/board/viewtopic.php?t=2696">here</a> and <a href="http://www.climateprediction.net/board/viewtopic.php?t=2745">here</a>.

It's a fascinating topic, and I agree that in at least a proportion of cases the team do seem to be matching processor types under Windows. None of us are quite sure what to make of it as yet; are the observed variations inherent in the model or the result of calculation error?

---

&gt; It's a fascinating topic, and I agree that in at least a proportion of cases
&gt; the team do seem to be matching processor types under Windows. None of us are
&gt; quite sure what to make of it as yet; are the observed variations inherent in
&gt; the model or the result of calculation error?

One reason is that we make an initial condition ensemble - we keep all the parametres the same but run the model from slightly different start fields. The other is that we also hand out duplicates: identical runs (right down to the starting conditions). Results suggest that the duplicates agree perfectly (as far as we've looked) most of the time, but can vary a bit. [It has to do with how different processors, math libraries, etc go about calculating the same equations.]
Basically, any small perturbation in a model field is going to drive the model away from an arbitrarily close model. It's the butterfly effect in action.

Dave

---

&gt; One reason is that we make an initial condition ensemble - we keep all the
&gt; parametres the same but run the model from slightly different start fields.
&gt; The other is that we also hand out duplicates: identical runs (right down to
&gt; the starting conditions). Results suggest that the duplicates agree perfectly
&gt; (as far as we've looked) most of the time, but can vary a bit. [It has to do
&gt; with how different processors, math libraries, etc go about calculating the
&gt; same equations.]
&gt; Basically, any small perturbation in a model field is going to drive the
&gt; model away from an arbitrarily close model. It's the butterfly effect in
&gt; action.
&gt;
&gt; Dave
&gt;
Hi Dave,

Thanks for this. When you say 'run the model from slightly different start fields', does every different set of start fields generate a different Dtheta which is shown on the varied parameters list for each model as 'initial condition parameter'? I have only seen 10 values for dtheta, 0 through to 0.09, though I have heard mentioned that you want an initial condition emsemble size of hundreds at least for a few models near the one with all default values. How many different sets of start fields are you using for each Dtheta value?

---

Visit BOINC WIKI for help

And join BOINC Synergy for all the news in one place.

&gt; Thanks for this. When you say 'run the model from slightly different start
&gt; fields', does every different set of start fields generate a different Dtheta
&gt; which is shown on the varied parameters list for each model as 'initial
&gt; condition parameter'? I have only seen 10 values for dtheta, 0 through to
&gt; 0.09, though I have heard mentioned that you want an initial condition
&gt; emsemble size of hundreds at least for a few models near the one with all
&gt; default values. How many different sets of start fields are you using for each
&gt; Dtheta value?

That's right. At the moment we have support for an ic ensemble of ten. We need to change this and it may be simple or it may be just a bit trickier. We want to have an ic ensemble of several hundred for the standard run. Things ground to a halt at Christmas, as expected, but we hope to get on with this quite quickly this year. I know Dave S is pretty keen on it getting out there. I need to talk to Jamie, Neil and Tolu in the next week or so about it. Thanks for reminding me. :-)

Dave

---

This is all really interesting, because it means that when (mostly on the classic forum) correspondents have discussed and shown graphs for the same models turning out differently on different machines, this doesn't necessarily mean that some people's machines are producing unreliable results. Instead, there are known mathematical reasons why results can vary. Which is good to know.

For those of us who don't understand the maths terminology, could someone please explain what different start fields are? For me, this is a new concept.

---

__________________________________________________

&gt; For those of us who don't understand the maths terminology, could someone
&gt; please explain what different start fields are? For me, this is a new concept.

When we send out the models, they already have an atmosphere. The atmosphere is in a particular state - the grid boxes already have temperatures, pressures, humidities, etc. That - the atmospheric state before the model starts crunching - is the start field or initial condition (or a variety of other synonyms).

Climate models are chaotic models in which small differences in model fields grow over time, so that two very similar model atmosphereric states on (say) 5 January 1812 will lead to two very different atmosphereric states by about 30 January 1812. In the 1960s Ed Lorenz (from MIT in the US) worked out not only that this occurs in the atmosphere, but also that it arises because arbitrarily small perturbations (in spatial scale and energy) can cascade upwards and affect arbitrarily large (spatial and energy) scales in finite time.

Because of this property of models we run initial condition ensembles (the same model physics but with different start files) to try to get some idea of the spread of model behaviours for that set of model physics. It is our way of trying (statistically) to account for/deal with the chaos in the system.

Dave

---

Thanks, Dave. I hope you are also going to copy your post onto the classic forum because it will help clarify the worries about why the same model can produce different results. It's obviously very important, as you've said before elsewhere, to be sure of the validity of the model results before using them to predict the future.

---

__________________________________________________

Of the four Host doing this WU, one gave error, the other doesn't have a tricle in many days. I guess it's just mine and another Host doing this WU :-(

---

Professor Desty Nova
Researching Karma the Hard Way