[Twelf-developer] Twelf Performance Comparison (ad hoc)

[Christopher Richards]

On Fri, 20 Dec 2002 23:15:06 -0500, Frank Pfenning wrote:

> Background: After a message by Chris Richards about the
> problems with SML/NJ, versions > 110.40 [hope I got that
> right]

SML/NJ 110.40 and post-110.40 have different respective problems.  The
former has a memory leak in the GC, which might account for some of
the numbers you posted.  The latter has no leak, but the work-around
for same makes the GC extremely slow.  (Loading the FPCC suite takes
almost 10 times as long.)

> MLton does not degrade at all, although at some point in the
> middle (#4) it does some expensive garbage collection (the added
> time is all in gc).
> 
> I don't really have an explanation for these numbers, but
> they strongly suggest Karl might benefit for working with
> MLton.

One thing to look out for here.  My experience running the FPCC suite
under MLton showed that MLton does not yield unused memory to the
system as well as SML/NJ (at all?).  For example, while SML/NJ memory
usage settled at ~200 MB at the end of a run, MLton usage settled at
~600 MB.  MLton's speed advantage may turn out moot if the extra VM
usage results in paging.

-- 
Chris