On Jan 15, 2006, at 5:20 AM, Ian Lynagh wrote:
On Sat, Jan 14, 2006 at 07:04:53PM -0800, Jason Dagit wrote:
I think the numbers speak for themselves.
I think they are saying different things to me and you.
Not as different as you think. But I should have said what I was
thinking. I can't expect people to read my mind :) I was just tired
at that point and needed a break...
If things have regressed then the regression should be fixed, not
consolidated by forcing the patch to remain in memory.
Well, my feeling is that we have an easy way to *help* the situation
(as my previous email showed, it doesn't make things worse), but also
darcs needs to work on only using a constant amount of space during
any operation. Maybe using fixed size buffers would be easier than
relying on lazy evaluation. But I dread implementing that.
You can find some of the discussion of the optimisation work from
before
in the archives, incidentally, e.g. around:
http://www.abridgegame.org/pipermail/darcs-devel/2005-April/
001872.html
http://www.abridgegame.org/pipermail/darcs-devel/2005-April/
001938.html
I'll read through those. Thanks for the links.
(I think all my earlier testing was with lots of small files rather
than
a single large one, so it's possible someone just needs to look at the
large file case. I can't think of any reason why that would be worse
OTTOMH, though, unless the patch parsing is being strict (which, if it
is the case, should again be fixed by the new patch format)).
I'm wondering if there is some bit of garbage that happens, that the
GC doesn't think can be reclaimed, whenever a line is parsed and it
doesn't add up to being a problem until you have many lines. If my
hypothesis is true all that extra garbage adds up to be just as much
as the original patch or more. In my test case I created a line that
was about 60 characters and then duplicated the line until the file
was big enough. I wonder if using one long line instead would give
better results.
And yes, I agree the patch format needs to be changed.
Thanks,
Jason
_______________________________________________
darcs-devel mailing list
darcs-devel@darcs.net
http://www.abridgegame.org/cgi-bin/mailman/listinfo/darcs-devel