Continuing to compare the performance of correct and buggy
versions of vector-map, and comparing that differential to
the effects of loop unrolling and other hand tuning, this
might be a good time to observe that all of those effects
are small compared to the differences between different
implementations of Scheme. To illustrate that, I ran the
various versions of vector-map, both correct and buggy,
in three different systems: one interpreted, one compiled,
and in compiled Larceny. Timings in seconds for the same
benchmarks as before, on the same Mac mini:
truth1 truth2 peak space gc time
(bytes) (truth1)
recursive8 interpreter 662.8 716.9 .9
recursive4 interpreter 514.4 613.8 .7
consing interpreter 377.4 575.5 1.4
recursive interpreter 317.8 349.2 1.3
consing compiler 234.3 249.0 100.7
buggy4 interpreter 213.0 342.1 .4
buggy interpreter 201.4 437.9 .3
buggy8 interpreter 200.8 306.9 .2
recursive compiler 114.1 151.2 24.7
recursive4 compiler 104.3 117.5 24.4
recursive8 compiler 103.0 114.6 24.8
buggy compiler 97.6 106.3 23.8
buggy4 compiler 95.3 106.5 23.8
buggy8 compiler 94.5 106.0 23.8
recursive Larceny 39.6 45.0 36000 .8
consing Larceny 36.0 40.4 12000 .7
recursive4 Larceny 27.5 28.3 18000 .4
buggy Larceny 24.6 29.2 4000 .2
recursive8 Larceny 23.7 27.8 11000 .2
buggy4 Larceny 21.9 26.7 4000 .1
buggy8 Larceny 19.2 22.5 4000 .1
bummed-rec8 Larceny 16.1 18.2 11000 .2
bummed-buggy8 Larceny 12.3 13.3 4000 .2
When performance matters, I would rather run correct code
on a fast system than buggy code on a slow system, but YMMV.
Will
--
To say that something suffers the disadvantages
of its advantages is a pretty minimal sort of irony.
-- John Cowan
Received on Mon Jun 25 2007 - 16:38:33 UTC