Highly variable run times within the pool of results for a project cannot be controlled. This seriously upsets the underlying statistics methodology used to compute credit per second computed for devices. We just have to live with it.

It's a sloppy shoulders from both fronts, developers in Berkeley who will say the projects have to generate regular runtime tasks, the project owners who will say it's unrealistic... non-deterministic computations cannot be told to run regular times... a needle-haystack search also takes an unpredictable long or short time to find the object.

It's fixable, it requires a ground up redesign, but who's going to spend costly [salaried] time on something which you can't sell on Ebay, AND, since the data is not public and WCG/IBM would not allow anyone near in a million years, no public effort can be undertaken to research and fix the matter.

Maybe the techs could put it on the agenda of the next annual BOINC workshop [though think no WCG representation did go to the last workshop], to design New_Credit2.