Hm. Here is a log-log chart of those results (sorry for html mail). I'm not really sure if log-log is the right tool to use for a O(nlog(n)) curve though.
I think the take-away is that this is outside the domain where any interesting break points occur. Maybe run more tests on the low end to find where the tapesort can generate a single tape and avoid the merge and see where the discontinuity is with quicksort for the various work_mem sizes. And can you calculate an estimate where the domain would be where multiple passes would be needed for this table at these work_mem sizes? Is it feasible to test around there? [image: Inline image 1] -- greg
