> This could be viewed as an opportunity. If HPPC or fastwhatever is a > superior device, we could *add* it under org.apache.mahout.coll, > leaving the colt stuff under math. Then we, over the long term, > improve the naming, and avoid and abrupt incompatibility. >
You could. Or you could add a maven dependency on fastutil (it is in maven central now) and just drop the development of independent collections package... Don't know which will appeal to the community more. We have been exchanging ideas with Sebastiano recently and after you get the same fast algorithms in place the difference in implementations don't seem to account for that much. I assume HPPC vs. fastutil will be a matter of taste more than speed in the near future (HPPC has open internals and a different architecture for containers/ iterators compared to the familiar j.u.c.). > Can someone suggest a way to get from Dawid's ascii graphics to a plan? > Copy-paste to an editor with fixed-width font? :) You can also click on those links I've sent, they should take you to an on-line Caliber benchmark page. Or run the benchmarks yourself. Sebastiano has made some speed improvements driven by this collaboration and fastutil maps are faster than when I compared them. Dawid
