This is the mail archive of the guile@sourceware.cygnus.com mailing list for the Guile project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]

Re: weird stats (aha!)



let me just repeat after Jim: BENCHMARKING IS WEIRD.

now, for this GC markbit business:

* the patch I sent you is kinda OK, except that it changes (by
  oversight) all startup GC constants, like the initial heap size
  etc.  which means that any comparisons of it with the stock Guile GC 
  are not to be trusted.

* so, I fixed the oversights and benchmarked things again.  results:

** for small heaps, the scheme with the mark bits in their own
   dedicated space loses by some 10-15%, due most probably to
   excessive indirections.

** for large heaps (i.e. large enough to cause swapping, it's all
   disgustingly configuration-dependent), there is a win, of some
   5-10% depending on the architecture and the amount of swapping.

* executive conclusions:

** I'm now stopping dicking around trying to reduce the pain of
   swapping without reducing swapping itself.

** the best way to reduce swapping seems to be implementing a
   generational marker and a lazy sweeper.  the latter is possible
   even now, and might be quite helpful.

** if anybody's interested in further tinkering, the patch and my
   benchmark results are available on request.

-- 
If all the economists in the world were laid end to end, it would probably
be a good thing.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]