[Starpu-devel] Equivalence vs substitutability

[Chris Hennick <christopherhe@trentu.ca>]

Hi all,

Suppose a business was running two different heuristic-optimization tasks, one to increase red-widget sales in France and one to increase blue-widget sales in Canada. Since an increase in profit would be just as valuable no matter which product it came from, I'd tend to treat these as different implementations of the same task despite their distinct output.

If optimizing blue-widget sales was CPU-intensive and optimizing red-widget sales was GPU-intensive, I'd want to use StarPU to schedule them, and I'd define an "iteration" as however long it took to find $X in extra profit. But in this case, the two implementations would probably be subject to separate diminishing returns, so each would have an upward trend in running time, and extrapolating that trend would give a better estimate of running time than using a long-term average.

Does StarPU have any way to estimate running times in this situation? I don't think I've seen any research on task scheduling where it was acknowledged that processes could be substitutable without being black-box equivalent; but it must be a common situation, since large businesses and governments have so many more subproblems than bottom lines.