Respond with benchmarks and an updated branch

[ikiwiki] / doc / todo / should_optimise_pagespecs.mdwn

I think there is a problem in my "dependency graph". As an example, 
[here](http://poivron.org/~nil/misc/ikiwiki_buggy_index) is the index 
ikiwiki generated for [my site](http://poivron.org/~nil/misc/ikiwiki_buggy_index)
(note that the site changed since this index was generated).

Some **HUGE** dependencies appear, clearly non optimal, like

    depends = A｜ B | A | C | A | D | A | E | A | F | A | G | ....

or 

    depends= A | B | C | D | A | B | C | D | A | B | C | D | ....

Couldn't isolate the cause, but some sources for this problem may be:

* related to the img module
* easily observable in my sire because one of my pages includes 80 resized images

Other special things in my templates and site:

* a sidebar with \[[!include pages="notes/\*" template=foo]] while notes.mdwn has 
  a \[[!include pages="notes/*"]] and uses the sidebar; removed it, doesn't change
* a template (biblio.tmpl) calling the "img" plugin with a template parameter as the
  image filename; removed it, doesn't change
* some strange games with tags whose page calls a "map" directive to show other tags
  shile tags are also used in tagclouds (in the sidebar and in the main pages)
* ...

I observed these problems (same *kind*, I didn't check in details) on
 
* ikiwiki 2.00gpa1 + v5.8.4 + Debian 3.1
* ikiwiki 2.3 + v5.8.8 + Ubuntu 7.04

I can think about reducung the size of my wiki source and making it available online for analysis.

-- NicolasLimare

> As long as these dependencies don't grow over time (ie, when a page is
> edited and nothing changed that should add a dependency), I wouldn't
> worry about them. There are many things that can cause non-optimal
> dependencies to be recorded. For one thing, if you inline something, ikiwiki
> creates a dependency like:
> 
> (PageSpec) or (file1 or file2 or file3 ...)
> 
> Where fileN are all the files that the PageSpec currently matches. (This
> is ncessary to detect when a currently inlined file is deleted, and know
> the inlining page needs an update.) Now consider what it does if you have
> a single page with two inline statements, that inline the same set of
> stuff twice:
> 
> ((PageSpec) or (file1 or file2 or file3 ...) or (PageSpec) or (file1 or file2 or file3 ...)
>
> Clearly non-optimal, indeed.
> 
> Ikiwiki doesn't bother to simplify complex PageSpecs
> because it's difficult to do, and because all they use is some disk
> space. Consider what ikiwiki uses these dependencies for.
> All it wants to know is: does the PageSpec for this page it's considering
> rebuilding match any of the pages that have changed? Determining this is
> a simple operation -- the PageSpec is converted to perl code. The perl
> code is run.
> 
> So the total impact of an ugly dependency like this is:
> 
> 1. Some extra data read/written to disk.
> 2. Some extra space in memory.
> 3. A bit more data for the PageSpec translation code to handle. But that
>    code is quite fast.
> 4. Typically one extra function call when the generated perl code is run.
>    Ie, when the expression on the left-hand side fails, which typically
>    happens after one (inexpensive) function call, it has to check
>    the identical expression on the right hand side.
> 
> So this is at best a wishlist todo item, not a bug. A PageSpec simplifier
> (or improved `pagespec_merge()` function) could be written and improve
> ikiwiki's memory and disk usage, but would it actually speed it up any?
> We'd have to see the code to the simplifier to know.
> 
> --[[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
>> amount of time. The optimize-depends branch in my git repository
>> avoids using `pagespec_merge()` for this (indeed it's no longer used
>> at all), and instead represents dependencies as a list of pagespecs
>> rather than a single pagespec. This does turn out to be faster, although
>> not as much as I'd like. --[[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 think I might have solved some of the problems mentioned there,
>>>> actually - `pagespec_merge` no longer needs to exist in my branch (although
>>>> 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
>>>>> future features like [[tracking_bugs_with_dependencies]]). --[[smcv]]

---

Some questions on your optimize-depends branch. --[[Joey]]

In saveindex it still or'd together the depends list, but the `{depends}`
field seems only useful for backwards compatability (ie, ikiwiki-transition
uses it still), and otherwise just bloats the index.

> If it's acceptable to declare that downgrading IkiWiki requires a complete
> rebuild, I'm happy with that. I'd prefer to keep the (simple form of the)
> transition done automatically during a load/save cycle, rather than
> requiring ikiwiki-transition to be run; we should probably say in NEWS
> that the performance increase won't fully apply until the next
> rebuild. --[[smcv]]

>> It is acceptable not to support downgrades.
>> I don't think we need a NEWS file update since any sort of refresh,
>> not just a full rebuild, will cause the indexdb to be loaded and saved,
>> enabling the optimisation. --[[Joey]]

>>> A refresh will load the current dependencies from `{depends}` and save
>>> them as-is as a one-element `{dependslist}`; only a rebuild will replace
>>> the single complex pagespec with a long list of simpler pagespecs.
>>> --[[smcv]]

Is an array the right data structure? `add_depends` has to loop through the
array to avoid dups, it would be better if a hash were used there. Since
inline (and other plugins) explicitly add all linked pages, each as a
separate item, the list can get rather long, and that single add_depends
loop has suddenly become O(N^2) to the number of pages, which is something
to avoid..

> I was also thinking about this (I've been playing with some stuff based on the
> `remove-pagespec-merge` branch).  A hash, by itself, is not optimal because
> the dependency list holds two things: page names and page specs.  The hash would
> work well for the page names, but you'll still need to iterate through the page specs.
> I was thinking of keeping a list and a hash.  You use the list for pagespecs
> and the hash for individual page names.  To make this work you need to adjust the
> API so it knows which you're adding.  -- [[Will]]

> I wasn't thinking about a lookup hash, just a dedup hash, FWIW.
> --[[Joey]]

>> I was under the impression from previous code review that you preferred
>> to represent unordered sets as lists, rather than hashes with dummy
>> values. If I was wrong, great, I'll fix that and it'll probably go
>> a bit faster. --[[smcv]]

>>> It depends, really. And it'd certianly make sense to benchmark such a
>>> change. --[[Joey]]

>>>> Benchmarked, below. --[[smcv]]

Also, since a lot of places are calling add_depends in a loop, it probably
makes sense to just make it accept a list of dependencies to add. It'll be
marginally faster, probably, and should allow for better optimisation
when adding a lot of depends at once.

> That'd be an API change; perhaps marginally faster, but I don't
> see how it would allow better optimisation if we're de-duplicating
> anyway? --[[smcv]]

>> Well, I was thinking that it might be sufficient to build a `%seen`
>> hash of dependencies inside `add_depends`, if the places that call
>> it lots were changed to just call it once. Of course the only way to
>> tell is benchmarking. --[[Joey]]

>>> It doesn't seem that it significantly affects performance either way.
>>> --[[smcv]]

In Render.pm, we now have a triply nested loop, which is a bit
scary for efficiency. It seems there should be a way to
rework this code so it can use the optimised `pagespec_match_list`,
and/or hoist some of the inner loop calculations (like the `pagename`)
out.

> I don't think the complexity is any greater than it was: I've just
> moved one level of "loop" out of the generated Perl, to be
> in visible code. I'll see whether some of it can be hoisted, though.
> --[[smcv]]

>> The call to `pagename` is the only part I can see that's clearly
>> run more often than before. That function is pretty inexpensive, but..
>> --[[Joey]]

>>> I don't see anything that can be hoisted without significant refactoring,
>>> actually. Beware that there are two pagename calls in the loop: one for
>>> `$f` (which is the page we might want to rebuild), and one for `$file`
>>> (which is the changed page that it might depend on). Note that I didn't
>>> choose those names!
>>>
>>> The three loops are over source files, their lists of dependency pagespecs,
>>> and files that might have changed. I see the following things we might be
>>> doing redundantly:
>>>
>>> * If `$file` is considered as a potential dependency for more than
>>>   one `$f`, we evaluate `pagename($file)` more than once. Potential fix:
>>>   cache them (this turns out to save about half a second on the docwiki,
>>>   see below).
>>> * If several pages depend on the same pagespec, we evaluate whether each
>>>   changed page matches that pagespec more than once: however, we do so
>>>   with a different location parameter every time, so repeated calls are,
>>>   in the general case, the only correct thing to do. Potential fix:
>>>   perhaps special-case "page x depends on page y and nothing else"
>>>   (i.e. globs that have no wildcards) into a separate hash? I haven't
>>>   done anything in this direction.
>>> * Any preparatory work done by pagespec_match (converting the pagespec
>>>   into Perl, mostly?) is done in the inner loop; switching to
>>>   pagespec_match_list (significant refactoring) saves more than half a
>>>   second on the docwiki.
>>>
>>> --[[smcv]]

Very good catch on img/meta using the wrong dependency; verified in the wild!
(I've cherry-picked those bug fixes.)

----

Benchmarking results: I benchmarked by altering docwiki.setup to switch off
verbose, running "make clean && ./Makefile.PL && make", and timing one rebuild
of the docwiki followed by three refreshes. Before each refresh I used
`touch plugins/*.mdwn` to have something significant to refresh.

I'm assuming that "user" CPU time is the important thing here (system time was
relatively small in all cases, up to 0.35 seconds per run).

master at the time of rebasing: 14.20s to rebuild, 10.04/12.07/14.01s to
refresh. I think you can see the bug clearly here - the pagespecs are getting
more complicated every time!

After the initial optimization: 14.27s to rebuild, 8.26/8.33/8.26 to refresh.
Success!

Not pre-joining dependencies actually took about ~0.2s more; I don't know why.
I'm worried that duplicates will just build up (again) in less simple cases,
though, so 0.2s is probably a small price to pay for that not happening (it
might well be experimental error, for that matter).

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.

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.

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]]

[[!tag wishlist patch patch/core]]