Note: The speed enhancements below are mostly about optimizing files and client-side stuff. You can also speed up your catalog on the server side with hardware and database optimization, though doing so is likely to require assistance from Innovative.
Faster-loading pages make users happier. Some speed-enhancing techniques also increase server performance and decrease the amount of bandwidth used. Page speed also factors into search engine rankings, though that is less of a concern for our catalog than our website.
Most of the improvements came from Google Page Speed recommendations.
Fact I found interesting: Before I applied the catalog changes, the directory contained 682 files. It now has 49. Having fewer files doesn’t inherently speed up the app, but it does speak to the nature of the changes.
There are two types of images in a Millennium “example set”: icons (or whole buttons as images) and media type indicators. A typical setup, until recently, had 145 or so button images. The most recent example set available from Innovative has 45 icons and a couple support images for making buttons. These images are coupled with some crufty markup to make the buttons work. The latest set also ships with 39 media indicators (in fairness, we’d use maybe 13 of those for our library).
CPL now has one image for all the icons and one image for all the media indicators. In total, the catalog screens have 8 images: one is required but never used, another never seen, and two others used infrequently. Yes, we grab book jackets from an external service, but that’s the extent of the catalog graphics.
An individual example icon weighs in between 600 bytes and 1.3K. Our icon sprite is 5.44K. It can be loaded once and cached on the user’s machine. For pages with many buttons, this results in far fewer HTTP requests and a smaller total payload. The same goes for media indicators.
How do we pull this off? We use two spriting techniques: ‘traditional’ sprites and pseudo sprites. The styles to makes these work can be found in the catalog stylesheet. What’s novel, and how you can pull this off in Millennium, is in carefully crafting what’re called ‘wwwoptions’:
# pseudo sprite example
ICON_BUT_REQUEST=<span class="but icon request">Request</span>
# traditional sprite example
IMAGE_MATTYPE3=/screens/spacer.gif" class="media-icon bookmp3
As you can see, in both cases activating sprites involves appending classes to a particular HTML element. In the case of the pseudo sprites, there’s a class that makes anything look like a button, one that pads the left-hand side and loads the icon sprite image into the space, and one (‘request’ above) that stipulates where in the sprite image the request icon appears. In the traditional sprite, the “IMAGE_MATTYPE3” option must be an image. Since there’s no getting around that, I used a spacer gif. It’s 43 bytes — the smallest practical image for the web. I then circumvented the wwwoption to add the classes to the image. Those classes set the background size and position the sprite image.
Be advised that pseudo sprites only work in standards-compliant browsers and IE8+. If a substantial number of your users are on IE7-, you may want to wait. Users with non-supporting browsers can still see the button style we use, but the icons won’t appear. Sites that work for everyone but look better for modern browsers are considered progressively-enhanced; that’s what you can consider these buttons.
Syndetics Plus, by default, loads onto a page something like this
- Page calls widget.js directly, via a script tag
- widget.js loads jQuery from Google’s CDN
- After jQuery is loaded, widget.js loads widget_connector.js
- widget_connector.js gets ready, finds ISBNs on the page, and requests an ISBN-specific widget_response.js
- widget_response.js writes to the page and asks widget_connector.js to do stuff
This is a mess. It’s a long chain of serialized resources that aren’t well-cached by the browser, aren’t served minified and/or gzipped, can’t be served over SSL, and that block full page rendering until they’re complete (whether or not there’s a result). Total payload for this process is 49K.
Librarything for Libraries is even worse (here I go again). It goes:
- widget.php loads connector.php
- connector.php pulls in connector_LB.css
- connector.php gets ready, finds ISBNs on the page, and requests an ISBN-specific widget_response.php
- widget_response.php writes to the page and asks connector.php to do stuff
CPL-specific code pulls the primary ISBN from bib record pages, validates it, then makes sure the page is served as regular HTTP. All of this occurs after the page has already loaded. If the conditions are met, we then feed the ISBN to the widget_response files from the two services, getting back just the results specific to the bib record. The widget_response files handle writing to the page all by themselves, so the service-specific code just adds things those scripts expect to find and handles certain on-page actions. Since we do this all asynchronously, the page loads (and feels like it loads) as quickly as Millennium allows.
What’s funny about the terribleness coming from Bowker’s servers is that a big part of improving the situation is both easy and a big sack of win. Users would get faster pages and Bowker’d save on bandwidth and hardware specs. Some quick apache config of mod_expires and mod_deflate is all that’s required to pay dividends. A deployment scheme involving minification and/or a refactor of the connector scripts would be more time-consuming, but could make the widgets substantially faster. Serving out static scripts without query strings would be a huge help, for instance.
I ran all the catalog images through PNGGauntlet, which uses PNGOUT to reduce file size — sometimes by 25% or more. For high-use files, I also weighed the quality reduction of making them 8-bit PNGs; the sprites, for instance, get served out as 8-bit files, making them substantially smaller.
I also rolled out a sprite for the site’s layout and theme, optimized the chat widget, and combined and minified CSS files for use by the catalog. Our web server, in contrast to the catalog, is pretty aggressive about caching. Files are served with far-future expires headers. APC for PHP caches the opcode to minimize file reads; this results in faster execution speed across the board, but not necessarily faster serving to the end user. We have drupal caching set up, which benefits from MySQL query caching. Apache gets some love from the OS’s buffer cache. Altogether, the website is a layer of caches.
I’d like to reduce the payload of every page on our site by rolling out a new theme. There’s a lot of extraneous markup and styling that can get cut out pretty easily. Other possibilities:
- Optimize images as they’re uploaded
- Defer jQuery and other scripts on page loads
- Put stuff into a cache manifest
- Investigate mod_pagespeed to do some of this stuff automatically
For the catalog, a few squeaks of extra speed could come from moving static files to our web server (which has longer cache lifetimes), combining the (minified) catalog CSS into the site CSS, and working with Syndetics to get book jackets from URLs without query parameters.