>
> --[[Joey]]
-[[!template id=gitbranch branch=smcv/ready/optimize-depends author="[[smcv]]"]]
-
>> I've been looking at optimizing ikiwiki for a site using
>> [[plugins/contrib/album]] (which produces a lot of pages) and it seems
>> that checking which pages depend on which pages does take a significant
>> rather than a single pagespec. This does turn out to be faster, although
>> not as much as I'd like. --[[smcv]]
+>>> [[Merged|done]] --[[smcv]]
+
>>> I just wanted to note that there is a whole long discussion of dependencies and pagespecs on the [[todo/tracking_bugs_with_dependencies]] page. -- [[Will]]
>>>> Yeah, I had a look at that (as the only other mention of `pagespec_merge`).
>>>> I haven't actually deleted it), because the "or" operation is now done in
>>>> the Perl code, rather than by merging pagespecs and translating. --[[smcv]]
-[[!template id=gitbranch branch=smcv/ready/remove-pagespec-merge author="[[smcv]]"]]
-
>>>>> I've now added a patch to the end of that branch that deletes
>>>>> `pagespec_merge` almost entirely (we do need to keep a copy around, in
>>>>> ikiwiki-transition, but that copy doesn't have to be optimal or support
refresh. I think you can see the bug clearly here - the pagespecs are getting
more complicated every time!
+> I can totally see a bug here, and it's one I didn't think existed. Ie,
+> I thought that after the first refresh, the pagespec should stabalize,
+> and what it stabalized to was probably unnecessarily long, but not
+> growing w/o bounds!
+>
+> a) Explains why ikiwiki.info has been so slow lately. Well that and some
+> other things that overloaded the system.
+> b) Suggests to me we will probably want to force a rebuild on upgrade
+> when fixing this (via the mechanism in the postinst).
+>
+> I've investigated why the pagespecs keep growing: When page A changes,
+> its old depends are cleared. Then
+> page B that inlines A gets rebuilt, and its old depends are also cleared.
+> But page B also inlines page C; which means C gets re-rendered. And this
+> happens w/o its old depends being cleared, so C's depends are doubled.
+> --[[Joey]]
+
After the initial optimization: 14.27s to rebuild, 8.26/8.33/8.26 to refresh.
Success!
though, so 0.2s is probably a small price to pay for that not happening (it
might well be experimental error, for that matter).
+> It's weird that the suggested optimisations to
+> `add_depends` had no effect. So, the commit message to
+> b6fcb1cb0ef27e5a63184440675d465fad652acf is actually wrong.. ? --[[Joey]]
+
+>> I'll try benchmarking again on the non-public wiki where I had the 4%
+>> speedup. The docwiki is so small that 4% is hard to measure... --[[smcv]]
+
Not saving {depends} to the index, using a hash instead of a list to
de-duplicate, and allowing add_depends to take an arrayref instead of a single
pagespec had no noticable positive or negative effect on this test.
+> I see e4cd168ebedd95585290c97ff42234344bfed46c is still in your branch
+> though. I don't like using an arrayref, it could just take `($page, @depends)`.
+> and I don't see the need to keep it if it doesn't currently help.
+
+>> I'll drop it. --[[smcv]]
+
+> Is there any reason to keep 7227c2debfeef94b35f7d81f42900aa01820caa3
+> if it doesn't improve speed?
+> --[[Joey]]
+
+>> I'll try benchmarking on a more complex wiki and see whether it has a
+>> positive or negative effect. It does avoid being O(n**2) in number of
+>> dependencies. --[[smcv]]
+
Memoizing the results of pagename brought the rebuild time down to 14.06s
and the refresh time down to 7.96/7.92/7.92, a significant win.
+> Ok, that seems safe to memoize. (It's a real function and it isn't
+> called with a great many inputs.) Why did you chose to memoize it
+> explicitly rather than adding it to the memoize list at the top?
+
+>> It does depend on global variables, so using Memoize seemed like asking for
+>> trouble. I suppose what I did is equivalent to Memoize though... --[[smcv]]
+
Refactoring to use pagespec_match_list looks more risky from a code churn
point of view; rebuild now takes 14.35s, but refresh is only 7.30/7.29/7.28,
another significant win.
--[[smcv]]
+> I had mostly convinced myself that
+> `pagespec_match_list` would not lead to a speed gain here. My reasoning
+> was that you want to stop after finding one match, while `pagespec_match_list`
+> checks all pages for matches. So what we're seeing is that
+> on a rebuild, `@changed` is all pages, and not short-circuiting leads
+> to unnecessary work. OTOH, on refresh, `@changed` is small and I suppose
+> `pagespec_match_list`'s other slight efficiencies win out somehow.
+>
+> Welcome to the "I made ikiwiki twice as fast
+> and all I got was this lousy git sha1sum" club BTW :-) --[[Joey]]
+
[[!tag wishlist patch patch/core]]
sipb-www / ikiwiki.git / blobdiff