ikiwiki/doc/tips/optimising_ikiwiki.mdwn

214 lines
8.2 KiB
Markdown

[[!meta date="2009-10-15 18:42:46 -0400"]]
Ikiwiki is a wiki compiler, which means that, unlike a traditional wiki,
all the work needed to display your wiki is done up front. Where you can
see it and get annoyed at it. In some ways, this is better than a wiki
where a page view means running a program to generate the page on the fly.
But enough excuses. If ikiwiki is taking too long to build your wiki,
let's fix that. Read on for some common problems that can be avoided to
make ikiwiki run quick.
[[!toc]]
(And if none of that helps, file a [[bug|bugs]]. One other great thing about
ikiwiki being a wiki compiler is that it's easy to provide a test case when
it's slow, and get the problem fixed!)
## rebuild vs refresh
Are you building your wiki by running a command like this?
ikiwiki --setup my.setup
If so, you're always telling ikiwiki to rebuild the entire site, from
scratch. But, ikiwiki is smart, it can incrementally update a site,
building only things affected by the changes you make. You just have to let
it do so:
ikiwiki --setup my.setup --refresh
Ikiwiki automatically uses an incremental refresh like this when handing
a web edit, or when run from a [[rcs]] post-commit hook. (If you've
configured the hook in the usual way.) Most people who have run into this
problem got in the habit of running `ikiwiki --setup my.setup` by hand
when their wiki was small, and found it got slower as they added pages.
## use the latest version
If your version of ikiwiki is not [[!version]], try upgrading. New
optimisations are frequently added to ikiwiki, some of them yielding
*enormous* speed increases.
## run ikiwiki in verbose mode
Try changing a page, and run ikiwiki with `-v` so it will tell you
everything it does to deal with that changed page. Take note of
which other pages are rebuilt, and which parts of the build take a long
time. This can help you zero in on individual pages that contain some of
the expensive things listed below.
## expensive inlines
Do you have an archive page for your blog that shows all posts,
using an inline that looks like this?
\[[!inline pages="blog/*" show=0]]
Or maybe you have some tag pages for your blog that show all tagged posts,
something like this?
\[[!inline pages="blog/* and tagged(foo)" show=0]]
These are expensive, because they have to be updated whenever you modify a
matching page. And, if there are a lot of pages, it generates a large html
file, which is a lot of work. And also large RSS/Atom files, which is even
more work!
To optimise the inline, consider enabling quick archive mode. Then the
inline will only need to be updated when new pages are added; no RSS
or Atom feeds will be built, and the generated html file will be much
smaller.
\[[!inline pages="blog/*" show=0 archive=yes quick=yes]]
\[[!inline pages="blog/* and link(tag)" show=0 archive=yes quick=yes]]
Only downsides: This won't show titles set by the [[ikiwiki/directive/meta]]
directive. And there's no RSS feed for users to use -- but if this page
is only for the archives or tag for your blog, users should be subscribing
to the blog's main page's RSS feed instead.
For the main blog page, the inline should only show the latest N posts,
which won't be a performance problem:
\[[!inline pages="blog/*" show=30]]
## expensive maps
Do you have a sitemap type page, that uses a map directive like this?
\[[!map pages="*" show=title]]
This is expensive because it has to be updated whenever a page is modified.
The resulting html file might get big and expensive to generate as you
keep adding pages.
First, consider removing the "show=title". Then the map will not show page
titles set by the [[ikiwiki/directive/meta]] directive -- but will also
only need to be generated when pages are added or removed, not for every
page change.
Consider limiting the map to only show the toplevel pages of your site,
like this:
\[[!map pages="* and !*/*" show=title]]
Or, alternatively, to drop from the map parts of the site that accumulate
lots of pages, like individual blog posts:
\[[!map pages="* and !blog/*" show=title]]
## sidebar issues
If you enable the [[plugins/sidebar]] plugin, be careful of what you put in
your sidebar. Any change that affects what is displayed by the sidebar
will require an update of *every* page in the wiki, since all pages include
the sidebar.
Putting an expensive map or inline in the sidebar is the most common cause
of problems. At its worst, it can result in any change to any page in the
wiki requiring every page to be rebuilt.
## avoid htmltidy
A few plugins do neat stuff, but slowly. Such plugins are tagged
[[plugins/type/slow]].
The worst offender is possibly [[plugins/htmltidy]]. This runs an external
`tidy` program on each page that is built, which is necessarily slow. So don't
use it unless you really need it; consider using the faster
[[plugins/htmlbalance]] instead.
## be careful of large linkmaps
[[plugins/Linkmap]] generates a cool map of links between pages, but
it does it using the `graphviz` program. And any changes to links between
pages on the map require an update. So, avoid using this to map a large number
of pages with frequently changing links. For example, using it to map
all the pages on a traditional, highly WikiLinked wiki, is asking for things
to be slow. But using it to map a few related pages is probably fine.
This site's own [[plugins/linkmap]] rarely slows it down, because it
only shows the index page, and the small set of pages that link to it.
That is accomplished as follows:
\[[!linkmap pages="index or (backlink(index)"]]
## overhead of the search plugin
Be aware that the [[plugins/search]] plugin has to update the search index
whenever any page is changed. This can slow things down somewhat.
## cgi overload workaround
If the ikiwiki.cgi takes a long time to run, it's possible
that under load, your site will end up with many
of them running, all waiting on some long-running thing,
like a site rebuild. This can prevent the web server from doing anything
else.
A workaround for this problem is to set `cgi_overload_delay` to
a number of seconds. Now if ikiwiki.cgi would block waiting
for something, it will instead display a Please wait message (configurable
via `cgi_overload_message`, which can contain arbitrary html),
and set the page to reload it after the configured number of seconds.
This takes very little load, as it all happens within compiled C code.
Note that it is currently limited to GET requests, not POST requests.
## scaling to large numbers of pages
Finally, let's think about how huge number of pages can affect ikiwiki.
* Every time it's run, ikiwiki has to scan your `srcdir` to find
new and changed pages. This is similar in speed to running the `find`
command. Obviously, more files will make it take longer.
You can avoid this scanning overhead, if you're using git, by setting
`only_committed_changes`. This makes ikiwiki --refresh query git for
changed files since the last time, which tends to be a lot faster.
However, it only works if all files in your wiki are committed to git
(or stored in the [[/plugins/transient]] underlay).
* Also, to see what pages match a [[ikiwiki/PageSpec]] like "blog/*", it has
to check if every page in the wiki matches. These checks are done quite
quickly, but still, lots more pages will make PageSpecs more expensive.
* The backlinks calculation has to consider every link on every page
in the wiki. (In practice, most pages only link to at most a few dozen
other pages, so this is not a `O(N^2)`, but closer to `O(N)`.)
* Ikiwiki also reads and writes an `index` file, which contains information
about each page, and so if you have a lot of pages, this file gets large,
and more time is spent on it. For a wiki with 2000 pages, this file
will run about 500 kb.
If your wiki will have 100 thousand files in it, you might start seeing
the above contribute to ikiwiki running slowly.
## profiling
If you have a repeatable change that ikiwiki takes a long time to build,
and none of the above help, the next thing to consider is profiling
ikiwiki.
The best way to do it is:
* Install [[!cpan Devel::NYTProf]]
* `PERL5OPT=-d:NYTProf`
* `export PERL5OPT`
* Now run ikiwiki as usual, and it will generate a `nytprof.out` file.
* Run `nytprofhtml` to generate html files.
* Those can be examined to see what parts of ikiwiki are being slow.