1. Web Caching and Content Delivery

2. Caching for a Better Web Performance is a major concern in the Web Proxy caching is the most widely used method to improve Web performance Duplicate requests to the same document served from cache Hits reduce latency, bandwidth demand, server load Misses increase latency (extra hops) ClientsProxy CacheServers Hits Misses Internet [Source: Geoff Voelker]

3. Proxy Caching How should we build caching systems for the Web? Seminal paper [Chankhunthod96] Proxy caches [Duska97] Akamai DNS interposition [Karger99] Cooperative caching [Tewari99, Fan98, Wolman99] Popularity distributions [Breslau99] Proxy filtering and transcoding [Fox et al] Consistency [Tewari,Cao et al] Replica placement for CDNs [et al] [Voelker]

4. Issues for Web Caching Binding clients to proxies, handling failover Manual configuration, router-based “transparent caching”, WPAD (Web Proxy Automatic Discovery) Proxy may confuse/obscure interactions between server and client. Consistency management At first approximation the Web is a wide-area read-only file service...but it is much more than that. caching responses vs. caching documents deltas [Mogul+Bala/Douglis/Misha/others@research.att.com] Prefetching, scale, request routing, scale, performance Web caching vs. content distribution (CDNs, e.g., Akamai)

5. End-to-End Content Delivery request stream Internet hosting network request distributor surrogate caches CDN servers proxies server array + storage upstreamdownstream

6. Proxy Cache Effectiveness How to measure Web cache effectiveness (goals)? Hit ratio Savings in bandwidth or server load Reduction in perceived user latency What factors determine/limit effectiveness? Capacity? User population? Proxy placement in the network? Updates and invalidations?

7. Web Traffic Characterization Research question: how do goals and traffic behavior shape strategies for deploying and managing proxy caches? Replacement policy: what objects to retain in cache? Large vs. small, relative importance of popularity and stability Deployment: where to place the cache? Close to server or client? How many users per cache? Prefetching? Since the Web is in active deployment on a large-scale, Web traffic characterization is an empirical science. Science of mass behavior: observe and test hypotheses.

8. Zipf [Breslau/Cao99] and others observed that Web accesses can be modeled using Zipf-like probability distributions. Rank objects by popularity: lower rank i ==> more popular. The probability that any given reference is to the ith most popular object is p i Not to be confused with p c, the percentage of cacheable objects. Zipf says: “p i is proportional to 1/i , for some  with 0 <  < 1 ”. Higher  gives more skew: popular objects are way popular. Lower  gives a more heavy-tailed distribution. In the Web,  ranges from 0.6 to 0.8 [Breslau/Cao99]. With  =0.8, 0.3% of the objects get 40% of requests.

9. Zipf-like Reference Distributions p i 1/i   p i = 1 Probability of access to the object with popularity rank i: (This is equivalent to a power-law or Pareto distribution.) such that: head tail [Zipf 49, Duska et al. 97, Breslau et al. 98] Popularity rank heavy tail pipi

10. Importance of Traffic Models Analytical models like this help us to predict cache hit ratios (object hit ratio or byte hit ratio). E.g., get object hit ratio as a function of size by integrating under segments of the Zipf curve …assuming perfect LFU replacement Must consider update rate Do object update rates correlate with popularity? Must consider object size How does size correlate with popularity? Must consider proxy cache population What is the probability of object sharing? Enables construction of synthetic load generators SURGE [Barford and Crovella 99]

11. The “Trickle-Down Effect” clients cache to servers floodtrickle What is the effect on “downstream” traffic? What is the significance of this effect? How does it impact design choices for components “behind” the caches?

12. A Look at the Miss Stream synthetic trace SURGE-generated low locality:  = 0.6 log-log plot head: flattened midrange: tapers tail: intact Zipf-like 1035 816

13. 1998 ibm.com high locality fit Zipf  = 0.76 skewed: 77 % / 1% Effect on Server Trace (ibm.com)

14. What’s Happening? (LRU) Suppose the cache fills up in R references. (That’s a property of the trace and the cache size.) Then a cache miss on object with rank i occurs only if i is referenced…. probability p i …and i has not been referenced in the last R requests. probability (1 - p i ) R Stack distance P(a miss is to object i) is q i = p i (1 - p i ) R

15. Miss Stream Probability by Popularity q i : R = 10 4,  =0.7 IBM 1998 (32 MB) Moderately popular objects now dominate.

16. Object Hit Ratio by Popularity (1) synthetic  = 0.6

17. Object Hit Ratio by Popularity (2) IBM 1998

18. Limitations/Features of This Study static (cacheable) objects ignore misses caused by updates invalidation/expiration LRU replacement vary cache effectiveness by capacity cache intercepts all client traffic ignore effect on downstream traffic volume

19. Proxy Deployment and Use Where to put it? How to direct user Web traffic through the proxy? Request redirection Much more to come on this topic… Must the server consent? Protected content Client identity “Transparent” caching and the end-to-end principle Must the client consent?

20. Interception Switches ISP cache array The client doesn’t know. The server doesn’t know. Neither side told HTTP to disable it. Is it legal? Good thing? Bad thing?

21. Shouldn’t This Be Illegal? end middle RFC 1122: The Internet Architecture (IPv4) specifies that each packet has a unique destination “host” address. Problems middle boxes may be subversive IPsec and SSL dynamic routing

22. Cache Effectiveness Previous work has shown that hit rate increases with population size [Duska et al. 97, Breslau et al. 98] However, single proxy caches have practical limits Load, network topology, organizational constraints One technique to scale the client population is to have proxy caches cooperate