Isn't this lengthy thesis about saying that doing a map reduce on consecutive data that have are hot and localized in memory is all the better that we use commutative/distributive operations. Well it was trivial.<p>Well, I would have prefered a first year student demonstrating the trivia that operation like sorting multidimensional vector will kick by the nature less optimization possible in CPU and GPU and that non linear operation requires reading the whole sample to have a non random quantifiable error.<p>Hence, map reduce performance is better for combination of distributive linear operation (like ARMA, and I still wonder about hilbertian geometry) and horrible for non linear mapped functions (median filter, nth percentile, sort, top X).