This tidbit popped up on LtU today, about performance differences computing large numbers (like 3^1,000,000).
It starts out with a usenet post that compares two Scheme implementations and the performance differentials manipulating very large numbers. Brad Lucier starts by highlighting some shortcomings with MzScheme's numeric handling, as compared to an as-yet-unreleased version of Gambit-C 4.0.
Brad observes that MzScheme's performance is slow in most places, abysmal in others. The nature of the deviations leads Brad to conclude that there's something very seriously wrong with bignum handling under the hood. A little while later, Matthew Flatt (one of the MzScheme developers) responds that Brad's experiments were indeed correct, and that MzScheme needed some tweakage to use gmp better.
The as-fast-as-it-should-be version of MzScheme is now available as a nightly build. The bugfixes will be incorporated into the next release.
