Saturday, July 7, 2007

The Problem with SlickSpeed

For the past month or so, there's been a lot of noise about the SlickSpeed Selectors Test Suite. Since I'm in the market for a good selector engine for Zillow, and since it's a bit of a rite of passage (a front-end web dev's equivalent of compiler authoring?), I wrote my own, to see how well I could do and to see how it stacks up to the rest.

So of course, I modified the suite (under its MIT license) to test my little attempt as well. I was pleased with my initial results, but found the test document that comes packaged with the suite to be a little simplistic. Not enough variety or depth of nesting; the resulting DOM structures don't really resemble what I look at on a daily basis at work. I wanted to measure performance in the wild. So I replaced Shakespeare's As You Like It with the Home Details Page for, perhaps the most complex page on the site. Among other things, it includes a photo gallery, a Flash map, an Live Local-based Bird's Eye View map, a chart widget, several ads, tables, etc.

The results, you can see for yourself, here.

As it turns out, according to SlickSpeed, my engine outperforms all but 2 of the other engines on Firefox, and is the best performer on IE7.

So my misgivings on the nature of the document wandered over to the construction of the queries. The given queries perform a "breadth" pass, but they don't really provide a "depth" pass including all manner of combinations of possible selectors, so I wrote my own addition to the suite that picks random elements from the DOM and generates a matching CSS1 selector for it. You can see the dynamic test suite here.

Now, my engine's performance is fair to decent at best, but no longer the front-runner. Unless I can iron out the kinks, I might look into Ext's engine, especially since it fits nicely into the Yahoo! UI library we use at Zillow.

On the other hand, my engine is stand-alone and does not provide any other services or dependencies. It's a whopping 6KB (minified), but I wouldn't recommend the use of a CSS query engine for anything short of a full-scale web application anyway.

Some thoughts, though: For reasons that should be self-explanatory, it appears that all of the CSS engine makers are optimizing for Firefox. And once again, Opera's JavaScript engine (named Linear B) and DOM implementation beats out all the rest. Performance on IE looks to be all-around poorer. The Internet Explorer Team certainly has their work cut out for them, not only in improving their DOM and JScript performance and their developer tools (a decent profiler and a debugger that's not attached to Visual Studio would be nice), but also in winning over a hostile developer community. I guess that's what happens when the maker of the World's Number One Browser shuts down their development team for 5 years.

Prototype and MooTools appear to be compiling the CSS selectors into XPath statements for Firefox's (and Opera's) built-in xpath evaluator (too bad IE forgot to allow MSXML's XPath engine to apply to the HTML DOM). While the DOM performance for these XPath-based implementations is fantastic, they also help underline the end-user experience difference between browsers. Let's hope users take notice how much faster the leading non-IE browsers are in comparison; it's hard to win users over on the basis of standards compliance alone.

If nothing else, I hope my modified SlickSpeeds will help CSS query engine developers focus on what's important: CSS1 queries. The time scores at the bottom of the SlickSpeed test skew heavily toward obscure pseudoclass and attribute selectors which I for one won't use most of the time. It's the meat-and-potatoes tag, class, and ID that really count.


Candid.Answers said...
This comment has been removed by a blog administrator.
David Golightly said...

If you have something to sell me, you can always talk to my spam mailbox on my gmail account. Otherwise, please keep comments reasonably topical. Thanks!

Anonymous said...

Hi David,
Sorry your post on my blog got caught by WordPress and I just found it today. I agree the markup used in the test is pretty bad and your test looks much more realistic. Thanks for sharing your results!

Anonymous said...

Your links are not working. Please take a look.