Web Caching Schemes For The Internet – cont. By Jia Wang.

Web Caching Schemes For The Internet – cont.

By Jia Wang

Topics

Cache resolution/routing Prefetching Cache replacement Cache coherency Other topics

Cache Resolution/Routing

Most Web caching schemes: many Web caches scattered over the Internet

Main challenge: how to quickly locate the appropriate cache

No necessary location among document’s cache location

Unmanageably large cache routing tables

Cache Resolution/Routing

Out-of-date cache routing information leads to cache miss

Minimize the cost of a cache miss: ideal cache routing algorithm

Cache Resolution/Routing

Common approach – caching distribution tree from popular servers towards high demand sources – Resolution via cache routing table/ hash

functions Works well for popular documents

Cache Resolution/Routing

What about less popular documents?

Cache Resolution/Routing

Hit rate on web caches < 50%

Cache Routing Table

Malpani – make a group of caches function as one– Cache is selected arbitrary– In case of miss: use IP multicast (why?)– Redirection

Cache Routing Table

Cache Routing Table

Advantages:– No bottlenecks

– No single point of failure

Cache Routing Table

Disadvantage:– Overhead

Solutions?

Cache Routing Table

Harvest – organize caches in hierarchy– Internet Cache protocol: ICP– In case of miss: siblings and upward

Cache Routing Table

Adaptive Web Caching – mesh of caches– Distribution trees are built– Overlapping multicast groups– No root node will be overloaded– For less popular objects: long journey

Hashing Function

Cache Array Routing Protocol: CARP– “query-less” caching by hash function– Based on “array membership list” and URL

for exact cache location– Proxy removal: reassign 1/n URLs and

distribute new hash function

Prefetching

Caching documents at proxies improve Web performance with limited benefit

Maximum cache hit rate < 50%

Prefetching

One way to increase hit rate: anticipate future requests and preload or prefetch

Prefetching

Prefetching must be effective (why?) Prefetching can be applied in 3 ways:

– Between browser clients and Web Servers– Between proxies and Web Servers– Between browser clients and proxies

Between browser clients and Web Servers Cunha – use a collection of Web clients

– How to predict user’s future Web accesses from his past Web accesses

– Two types of users: net surfer and conservative

Between browser clients and Web Servers Conservative – easy to guess which

document will access next– Prefetching is well paid off

Net surfer – all documents have equal probability of being accessed– Price to be paid in terms of extra

bandwidth is too high

Between proxies and Web Servers Markatos –

– Web servers push popular documents to Web proxies (top-10)

– Web proxies push popular documents to Web clients

– Web servers can anticipate > 40% client’s request

– Requires cooperation from Web servers

Between proxies and Web Servers Performance:

– Top-10 manages to prefetch (up to) 60% of future requests

– less than 20% corresponding increase in traffic

Between browser clients and proxies Fan – reduce latency by prefetching

between caching proxies and browsers– Relies on the proxy – predict which cached

documents a user might reference next– Use idle time between user requests to

push documents to the user– Reduce client latency by 23%

Prefetching - summary

First two approaches – increase WAN traffic

Last approach – affects traffic over modems/LANs

Cache placement/replacement

Document placement/replacement algorithm can yield high hit rate

Cache placement – not been well studied

Cache replacement – can be classified into 3 categories:

Cache placement/replacement

Traditional policies

Key-based policies

Cost-based policies

Cache replacement – traditional policies Least Recently Used

– LRU Least Frequently Used

– LFU Pitkow/Recker – LRU

except if all objects are accessed within the same day

Cache replacement – key-based policies Size – evicts the

largest object (why?)

LRU-MIN – biased in favor the smaller objects– Evicts the LRU object which has size > S,

S/2, S/4 etc.

Cache replacement – key-based policies LRU-Threshold – LRU but objects

which have size > Threshold are never cached

Lowest Latency First

Cache replacement – cost-based policies GreedyDual-Size – associates a cost

with each object– Evicts object with lowest cost/size

Server-assisted – models the value of caching an object in terms of its fetching cost, size and cache prices– Evicts object with lowest value

Cache coherency

Caches provides lower access latency Side effect: stale pages Every Web cache must update pages in

its cache

Cache coherency

HTTP commands that assist Web proxies in maintaining cache coherence:

HTTP GET Conditional GET: HTTP GET combined

with the header IF-Modified-Since Pragma:no-cache Last-Modified:date

Cache coherence mechanisms

Current cache coherency schemes provides two types of consistency– Strong cache

consistency

– Weak cache consistency

Strong cache consistency

Client validation – polling every time – Cached resources are potentially out-of-

date– If-Modified-Since with each access to

proxy– Many 304 responses

Strong cache consistency

Server invalidation – Upon detecting a resource change, send

invalidation message– Server must keep track

of lists of clients– The lists can become

out-of-date

Weak cache consistency

Piggyback invalidation: three invalidation mechanisms are proposed:

Weak cache consistency

The Piggyback Cache Validation: on every communication between proxy to server, the proxy piggybacks a list of cached resources for validation

Weak cache consistency

The Piggyback Server Invalidation: on every communication between server to proxy, the server piggybacks a list of resources that have changed since the last access

Weak cache consistency

Combination of PSI and PCV: depends on the time since the proxy last requested invalidation

– If the time is small: PSI– For longer gaps: PCV

More topics

Load balancing– Hot-spot

Dynamic data caching– Active cache

Conclusion

Web service becomes more popular– More network congestion– More server overloading

Web caching – one of the effective techniques

Open problems – proxy placement, cache routing, dynamic data caching, fault tolerance, security etc.