1. Evaluation of Delivery Techniques for Dynamic Web Content Mor Naaman, Hector Garcia-Molina, Andreas Paepcke Department of Computer Science Stanford University {mor, hector, paepcke}@cs.stanford.edu http://www-db.stanford.edu/

2. Dynamic Web is Ubiquitous

3. Problems with Dynamic Pages Generation of pages is resource-intensive Pages are too dynamic, or too personalized, to be cached Higher load on servers (page generation and delivery) More network traffic

4. We Evaluate Two Competing Solutions (Both address at least the network load) ESI (Oracle, Akamai) Enables assembly of pages from small fragments Fragments can be cached on specialized network caches (edge servers) Fragments are assembled on the edge server Class Based Delta Encoding Computes delta of generated page from a chosen base file Base files can be cached on network caches Client receives delta from the server and base file from cache; applies delta to base file to get final page

5. A Page Content Model Page composed from groups; groups include items. Page construction modeled as two-phase selection (groups, then items) Groups Items

6. Our Simulation Book pages in Amazon-style website MyYahoo-type personalized pages Personalized stock portfolio pages A simple personalized weather page Test-case web pages:

7. Simulation of ESI Assuming Zipf-like distribution for groups and items (popularity i =k/i  ) Performance highly dependant on  (ranging from 0.7-1.5 in our simulations) Hit rate estimates for items: =Arrival rate; TTL = item time-to-live; = constant Sample simulation results (bookstore-type resource, With “backend” servers) Traffic vs. TTL Hit-rate vs. value of Zipfian parameter

8. Class-Based Delta Encoding Simulation For some pages, client likely to be able to re-use base files Traffic vs. number of base files Traffic vs. Same-Base threshold For other pages, client-cache link traffic is higher than before. To minimize client traffic, use same base file owned by client if delta is larger than threshold

9. Sample Comparison Numbers MyYahoo-type pages Amazon-style Book pages Savings – client link Savings – server link Edge cache usage ESI0%62%1.5Mb DE66%87%3.2Mb Savings – client link Savings – server link Edge cache usage ESI0%30%1.2Mb DE-8%82%2.2Mb

10. Conclusions Excellent *, Good +, Bad -, Sometimes ~ All the details: http://dbpubs.stanford.edu/pub/2003-7 ESIDE Reduces server traffic+* Reduces client traffic-~ Reduces computational load on web server*- Performance dependent on web page structureYes Performance dependent on characteristics of dataYesNo Benefits greater when popularity risesYesLess Requires main site hardware/software installationNoYes Requires web-page code changesYesNo Requires network infrastructure (CDN services)YesNo Can exploit information available from CDN for page construction YesNo