Cooperative Caching in the Internet

1
Cooperative Caching in the Internet
Presented by Mohammad Salimullah Raunak Feb. 15,
2000
2
Outline

• Motivation
• General Idea
• Papers
• Harvest and Squid (http//catarina.usc.edu/danzig/
  cache.ps)(http//www.ircache.net/7ewessels/Paper
  s/icp-squid.ps)
  
• Summary Cache (http//www.cs.wisc.edu/cao/papers
  /summarycache.ps)
  
• Cache Digests (http//www.ircache.net/7ewessels/
  Papers/wcw98.dvi)
  
• Crisp (http//www.research.att.com/misha/crisp/d
  istrProxy/hotos.ps)
  
• Summary

3
Motivation

• Proxy Caching
• Saves Bandwidth
• Reduces Latency
• Problems with a Single Proxy
• Bottleneck
• Scalability
• Single Point of Failure

4
General Idea
5
Harvest

• Hierarchical, Tree-like
• Query-response
• UDP based
• Twice as fast as CERN object cache
• Ten times faster than Netscape's Netsite.
• Evolved into
• Squid
• NetCache

6
ICP

• Internet Cache Protocol is the cache to cache
  communication Protocol developed in Harvest
  
• A lot of commercial and research cache systems
  have implemented ICP
  
• ICP is based on UCB to expedite query/reply
  exchange messages
  
• "ICP loss" provides rudimentary load balancing
  mechanism
  
• Suffers from the security problems of UDP less
  effective when cache servers are not close to
  each other

7
Summary Cache

• Individual querying with ICP increases overhead
• Inter-proxy traffic increases by a factor or 70
  to 90
  
• In Summary Cache
• A cache directory is maintained
• Each proxy keeps a compact summary of the
  directory
  
• Probe all the summaries and do a targetted query
• The summaries are not 100 correct
• Issues
• Representation of the summary
• Frequency of the Summary Update

8
Summary Cache Issues

• Representation of the summary
• Objective Less Memory
• Choices
• Exact directory with MD5 signature
• Server Name
• Get something less memory and minimal false hits
• One Solution
• Use Bloom Filters

9
Bloom Filters

• A hash based probabilistic scheme that answers
  membership queries
  
• Zero probability for false negatives
• low probability for false positives
• A vector v of m bits (initially set to 0)
• k independent hash functions, h1, h2, ... hk
• For each a in A , the bits at positions h1(a),
  h2(a), ... hk(a) in v are set to 1.

10
Bloom Filters (Contd.)

• For a query for object b in A , check the bits at
  positions h1(b), h2(b), ... hk(b) in v are set
  to 1
  
• If any of them is 0, b is not in A
• Otherwise, assume b is in the set with certain
  probability of hitting a false positive
  
• Each proxy maintains a local Bloom filter to
  represent its own cached documents.
  
• For each location in the bit vector, it also
  keeps a counter to keep track of how many times
  the bit was set.

11
Bloom Filters (Contd.)

• A proxy builds a Bloom filter from the list of
  URLs of its cached documents
  
• It sends the bit array plus the specification fo
  the hash function to other proxies
  
• Update can happen by sending the whole bit array
  or the changes.
  
• A balanced tradeoff between the memory
  requirement and the false positive ratio

12
Summary Cache Issues

• When to update?
• Impact of delayed update
• Update the summaries only when the percentage of
  new documents has reached a threshold
  
• Use false misses as an indicator
• A delay threshold of 1 to 10 for updating
  summaries results in a tolerable degradation of
  the cache hit ratios
  
• Use either broadcast or periodic exchange for
  updating summaries

13
Cache Digests

• Rousskov and Wessels at NLANR
• Very very similar to Summary Cache
• Digests allow proxies to make information about
  their cache contents available to peers in a
  compact form
  
• A proxy uses digest to identify the likely
  neighbours to have the missed object
  
• Digests are also based on Bloom Filters
• While ICP generates a steady stream of small
  packets, with Bloom filter based approaches,
  transfers occur in high volume bursts

14
Summary vs. Digests

• Same design objective
• Same technique (Bloom filters)
• Same update time policy
• Differrent data dissemination policy
• Summary cache uses push for updates
• Cache Digests uses pull based approach
• Cache Digests proposes to "piggyback" the update
  messages in HTTP replies.

15
CRISP

• Caching and Replication for Internet Service
  Performance
  
• A central directory of cached objects
• Proxies share their caches using this central
  mapping service
  
• Proxies notify the mapping service any time they
  add or remove and object from the cache.
  
• The updates and probing are done using unicast
  messages.

16
CRISP (cont.)
17
CRISP (cont.)

• Obvious Drawbacks
• Centralized structure
• Scalability problem
• Single point of failur
• CRISP argues
• Requests to the map is small in size many can be
  served
  
• The response time of a well-configured mapping
  server is well below the ability of a human
  percenption
  
• When needed, more mapping servers an be added and
  the URLs can be statically partitioned across the
  servers.
  
• Configure mapping server way below the saturation
  point

18
CRISP (cont.)

• Handling Failures
• When the mapping server fails, the cooperation
  becomes unavailable
  
• Proxies serve their clients in the usual way
• When a proxy fails, the mapping server notices it
  eventually and mards entries from that cache as
  unavailable
  
• When a proxy rejoins, it registers itself and its
  cache objects are restored or re-enabled in the
  directory
  
• When a mapping server comes alive from failure,
  the central map is recreated by joining the caches

19
Summary

• Cooperative caching improves performance
• Low overhead yet highly accurate information
  sharing mechanism is the prime objective
  
• Different approaches have been tried