I've previously mentioned the work Steve Souders is doing evangelizing
high performance websites at Yahoo! and I was very pleased to be able to hear him speak at the
FOWA conference. Sadly the video from that conference isn't going to be freely available and although the audio and slides will be freely available they are not yet available.
Luckily he has previously recorded a video of a very similar (though not identical) presentation and I've embedded it below.
It's 37 minutes so for those in a hurry here are my notes from his presentation at FOWA.
The key points have also been discussed in a series of blog posts so where possible I'll link out to the relevant posts.
Importance of the backend
The first point raised is that the
user perception of load time is more important than the actual load time. This means that the relevant metric is not how fast can the html document be returned to the browser but how quickly that html is rendered in the browser. In measuring how quickly a page renders it was quickly realised that the backend performance, returning the html for the page, accounted for only about
5% of the overall time it took to render the page. Even with a full cache the backend was still only 13% of overall time.
Of the top 10 sites in the US only the backend for Google accounted for more than 20% of the load time with a full cache. The Google homepage is so spartan that with a primed cache only two HTTP requests need to be made.
Cache Usage
The importance of the cache was then discussed with data presented discussing
how many people at a site had a primed cache. They inserted a one pixel image on the Yahoo! homepage and then monitored the number of HTTP requests with a 200 header (empty cache) and a 304 header (primed header).
It was found that 50% of daily users have an empty cache which accounts for 20% of daily pageviews. This varies depending on the type of site, for example an empty cache accounts for fewer pageviews in a webmail site where each user will view multiple pages, but is broadly accurate. The data highlights the importance of catering for those users without a primed cache. Excessive use of images can't be justified by the assumption that once they are loaded the cached versions can be used. 50% of your users every day will be arriving at your site with an empty cache.
iFrame
Next he talked briefly about iFrames and how they can cause a 40-50 ms delay. onLoad doesn't work until the iFrame source responds which can cause a problem with 3rd party content.
YSlow
Next he discussed YSlow which grades a website based on the
14 rules developed through their research. YSlow is an extension for Firebug the popular development extension for the Firefox browser. It looks at how the page was built. Despite looking at the content rather than the response time its score correlates well with the rendering time. As such it could be a valuable tool during development to predict the speed of a site prior to its launch.
Another issue which YSlow apparently solves is a bug in how Firebug charts HTTP requests. Apparently Firebug will show queries to the cache as HTTP requests and YSlow patches this.
That's all I made notes on. I've got vague memories of
stepping HTTP requests to increase download speed and
cookies are always worth considering but I picked up a nasty cold in London and it's all a bit fuzzy.
