Uploaded byJustin Dorfman
Predicting user activity to make the web fast presentation
This document discusses techniques for predicting user activity and prerendering web pages to improve page load times. It notes current median page load times of 2.3-4.3 seconds and proposes using idle time between user actions like typing a URL to prefetch and render pages the user is likely to navigate to next based on their browsing history and habits. The document outlines Google Chrome's implementation of this using the omnibox search suggestions to predict and prerender likely future pages, showing how it has resulted in over 15% of omnibox navigations being nearly instant by preloading pages. It encourages web developers to ensure their sites can support prerendering.
Predicting user activity to make the web fast presentation
- 1. Predicting User Activity to Make The Web Fast Velocity Conference 2012 Arvind Jain and Dominic Hamon Google
- 2. How fast isthe web today? ● Chrome ~ 2.3s/5.4s page load time (median/mean) ● Google Analytics ~ 2.9s/6.9s page load time (median/mean) ● Mobile ~ 4.3s/10.2s page load time (median/mean) http://analytics.blogspot.com/2012/04/global-site-speed-overview-how-fast-are.html
- 3. Web page sizeover time http://httparchive.org/
- 4. Making the webfast ● Faster browsers ● Faster networks ● Faster hardware ● Faster pages But not enough to make the web Instant...
- 5. The time betweenrequests ● ~ 3 seconds to type a URL ● ~ 15 seconds to select a search result ● Plenty of idle time ● Why not use it?
- 6. Predict and Prerender ●Predict user navigation or follow advice from page <link rel='prerender'> ● Fetch all resources for a page ● Render to a hidden tab ● Swap in when user navigates
- 7. Chrome Omnibox Prerendering ● Already provide suggestions to the user ● Opportunity to learn browsing behaviour ● Ability to develop good prediction model ● Little contention for resources
- 8. Implementation details ● Trackwhether the user takes a suggestion ● Map user's text input to suggestion and hit/miss counts ● Given user input and suggestion, calculate confidence C: ○ C = H / (H + M) where H is the hit count, M is the miss count
- 9. Implementation example ● Usertypes 'c' ● Suggestions are: ○ www.cnn.com ○ comcast.net ● User types 'n' ● Omnibox shows 'cn' ● Suggestion is www.cnn.com ● User selects www.cnn.com
- 10. Implementation example (cont.) ●We store: { 'c', { { 'cnn.com', 1, 0 }, { 'comcast.net', 0, 1 }, }, 'cn', { { 'cnn.com', 1, 0 }, }, }
- 11. Implementation example (cont.) ●Over time, this data structure evolves to something like: View your data at chrome://predictors in Chrome 20
- 12. Implementation example (cont..) ●User types 'c' ● Suggestions are: ○ www.cnn.com ○ comcast.net ● www.cnn.com is selected by default ● Confidence is C = 1 ● Start prerendering www.cnn.com
- 13.
- 14. Key results ● Coverage ○ Almost a third of Omnibox navigations are prerendered ● Accuracy ○ About 90% of those prerenders are used ● Instant ○ Between 15% and 20% of Omnibox navigations are instant (<10 ms) ● Median time saved ○ ~1 second (>40%) per Omnibox navigation
- 15. Total time savedper day Omnibox: Over 10 years Search: Over 20 years
- 16. Why should youcare? ● Your site will be prerendered at some point by some users ● Check your site is compatible ● Page Visibility API ● Consider <link rel='prerender'> for your own site https://developers.google.com/chrome/whitepapers/prerender http://prerender-test.appspot.com/
- 17. Thank you Arvind Jainand Dominic Hamon Google