Rosetta/BOINC does not validate against partial results. It should.

The typical Rosetta task runs multiple decoys (each of which I believe is an *independent* simulation). I had such a task terminate because while calculating decoy 7 came it up with a NAN. The results from the correctly completed previous 6 decoys were discarded.

Looked in the 'Workunit Details' page and saw that another system was identified as successfully completing that same task. The catch -- it did only 5 decoys.

There is something fundamentally unfair when ALL the work from a system that did more crunching gets discarded, while accepting work from a system that crunched less.
.

Argh! Thanks for the correction Mod.Sense. That "always under 3 hours" should be "almost always under 3 hours". I really should have caught that as my older ppc mac probably runs over its preferred runtime more than most.(Yep, just checked, my DCF is 1.18) I was thinking about this thread as I took a walk early this morning and by the time I got to the keyboard I was pretty focused on explaining why the "to completion" time doesn't match the estimate given by the project. On most if not all other projects this estimate is invisible to the cruncher and it's appearance/use here on Rosetta as the "target CPU run time" seems to lead to a fair amount of confusion. If you know about BOINC but not Rosetta or vice versa you could easily think something was wrong when in fact everything was operating just as it should.

WCG has some projects that run similar types of tasks in that they run as many models as they can within a set period of time. WCG has chosen 8 as the magic number and so far I have never seen one of those tasks run less than eight hours or more than a minute or so over eight hours. So either the runtime for each model is very very very short or the app calls for a finish at the eight hour mark regardless of the progress of the last model. I couldn't swear to it but I would bet on the latter explanation. If BOINC thinks the task will take eight hours and it takes eight hours the DCF will remain 1 and the "to completion" time (showing in the BOINC manager) for the next task will match the estimate (invisible to the cruncher) given to BOINC by the app. The "CPU time" and "to completion" columns will always(ur, maybe I should say almost always!) add up to eight and the percent in the "progress" column will (should) always make sense. But the trade off is the wasted cpu time spent working on that last model before being forced to abandon it by a strict task time limit. Given the many different types of tasks here on Rosetta, and the considerable variation in runtime per model I would guess that a strict task time limit here would result in a lot of wasted cpu time. Personally I prefer coping with the variation.


Thanks for the info on checkpoints. I've always been scared of poking about the BOINC files but I think I'd like to try this. Give me a week or so, to search through several forums for everything that could go wrong and then to build up my courage to try it anyway:)

Snags

Forgive my lack of knowledge, but is the "Target CPU Runtime" preference the option that changes how much work your client will retrieve? I just started running boinc with any sort of regularity lately. Thanks. :)