A couple of days ago I was excited to see my first Mid-Holocene model downloaded. Then I noticed the same job has been already assigned to another computer nearly a month ago, where it is about to be completed (phase 4 now). I started wondering if it would make much sense for me to do the same work but realized that maybe I could verify the other person’s findings. Meanwhile, the same model has been assigned to yet another person and, since his system is faster than mine, he will probably beat me to the punch too. http://climateapps2.oucs.ox.ac.uk/cpdnboinc/workunit.php?wuid=6535949

So, considering that it will take nearly a month of nonstop crunching on my system to do this job, and by that time it will be completed at least once (likely twice), isn’t my computer going to be more useful working on another model?

I realize that I will be getting the full credits either way, but credits are really not what I am after in this project. I would like to know that my computer is doing a valuable scientific research and not just replicating known data.

So what do you guys think I should do? Should I go for it or abort it and get another model?

---

Please continue to run it. Multiple results for the same model are part of the experiment. You could also look out for slow processing iceworlds that affect some model & processor/OS combinations: iceworld appeal.

---

Thanks for your advice. Does that mean some values are randomized on each run of the same model, so that each run is unique?

---

qwer6

You\'re not the first person to ask about this, and probably won\'t be the last.

I made a fairly lengthy reply to one person about this here on our other board.
Note the research paper mentioned near the bottom. If you want to hear about this straight from the project research people, this is where to find all the details.

This is a direct link to the paper.

---

Backups: Here

As far as I know all the models from the same workunit have the same starting values. But Tolu (CPDN\'s chief programmer) has said definitively that he expects almost every model from the same workunit to turn out slightly differently. This is because of tiny differences on individual machines. In any case I believe that Intel and AMD carry out the calculations slightly differently. I imagine that the longer the model the greater the differences.

Several years ago a small number of moderators contrived to run the same model from the same WU. We compared the graphs. They were similar but different.

The only models likely to be refused by CPDN quality control are the tiny proportion run on unstably overclocked computers. The value of the other accepted models was studied for an early CPDN publication and found to be just as good as models run on supercomputers.

---

Cpdn news

I can see how processors of different brands can produce slightly different results, due to special algorithms they use. However, as far as crunching on similar processors goes, there should not be any difference, unless one of them is overclocked and produces errors.

Speaking of which, let me go back to my Mid-Holocene model in question. http://climateapps2.oucs.ox.ac.uk/cpdnboinc/workunit.php?wuid=6535949. The first machine already completed the run, while mine and another one – apparently heavily overclocked - have phase one completed. If you care to compare the three phase 1 temp and precipitation graphs (all 3 cpus are Intel core2duo), you can clearly see how two of the graphs are identical, while the third one, produced by the oveclocked machine, differs. As far as I can tell the overclocked machine miscalculates and it does more harm than good to this project, since it takes a third machine to verify the results of the other two.

Overclocking while doing DC - not a good idea..

---

If you care to compare the three phase 1 temp and precipitation graphs (all 3 cpus are Intel core2duo), you can clearly see how two of the graphs are identical, while the third one, produced by the oveclocked machine, differs. As far as I can tell the overclocked machine miscalculates and it does more harm than good to this project, since it takes a third machine to verify the results of the other two.

Overclocking while doing DC - not a good idea..

The third one is not overclocked, simply running Linux while the other two are running Windows. The compiled version of hadsm3 type models uses higher optimizations on Linux than Windows. For the hadcm3, the opposite is true. The tiny differences you see in the first phase graphs is involved with different weather, not climate, i.e. if you look at timestep 254004 (or whatever) of phase 1 on the Windows vs. Linux PCs, the weather/temperature will be different between the two, but the global mean temperature will be extremely close. Since we\'re predicting and comparing climates from the models, this is what is important, not that timestep by timestep, they produce exactly the same output.

A paper was written on this previously by some of the researchers here.