CMP 788 Distributed IR

1
CMP 788 Distributed IR

• Part 2 Lecture 8
• Harvest- Part 3
• Harvest Replication and Object Caching
• Fall 05
• Department of Mathematics
• and Computer Science
• Lehman College, CUNY

2
Broker
3
SOIF Example
4
Replicator and Object Cache

• The Harvest Replicator can be used to replicate
  servers, to enhance user-base scalability.
• For example, the HSR will likely become heavily
  replicated, since it acts a point of first
  contact for searches and new server deployment
  efforts.
• The Replication subsystem can also be used to
  divide the gathering process among many servers
  (e.g., letting one server index each U.S.
  regional network), distributing the partial
  updates among the replicas.
• The Harvest Object Cache reduces network load,
  server load, and response latency when accessing
  located information objects.
• LFU (Least Frequently Used) cache replacement
  strategy is often used.
• Hierarchical caching if target object is not in
  the local cache, use subnets cache (often
  provided by firewall software), or parent cache
  (larger caches stored on server shared by many
  machines)

5
Hierarchical Cache Arrangement
6
Caching Subsystem
7
Caching Resolution Protocol

• Each cache in the hierarchy independently decides
  whether to fetch the reference from the objects
  home site or from its parent or sibling caches,
  using a simple resolution protocol.
• If the URL contains any of a configurable list of
  substrings, then the object is fetched directly
  from the objects home, rather than through the
  cache hierarchy.
• This feature is used to resolve non-cacheable
  (e.g., cgi-bin, password protected objects) URLs.
• If the URLs domain name matches a configurable
  list of substrings, then the object is resolved
  through the particular parent bound to that
  domain.
• If a cache receives a request for a URL that
  misses, it performs a remote procedure call to
  all of its siblings and parents, checking if the
  URL hits any sibling or parent.
• The cache retrieves the object from the site with
  the lowest measured latency.

8
Caching Resolution Protocol (contd)

• Hierarchies as deep as three caches add little
  noticeable access latency.
• The only case where the cache adds noticeable
  latency is when one of its parents fail, but the
  child cache has not yet detected it.
• In this case, references to this object are
  delayed by two seconds, the parent-to-child cache
  timeout.
• As the hierarchy deepens, the root caches become
  responsible for more and more clients.
• To keep root caches servers from becoming
  overloaded, Harvest hierarchy terminates at the
  first place in the regional or backbone network
  where bandwidth is plentiful. 

9
Caching Resolution Protocol (contd)

• Additionally, a cache option can be enabled that
  tricks the referenced URLs home site.
• This option allows the cache to retrieve the
  object from the home site if it happens to be
  closer than any of the sibling or parent caches.
• Can be based on estimating object access latency
  time using network ping or echo.
• A cache resolves a reference through the first
  sibling, parent, or home site to return a UDP
  Hit packet (through echo port).
• The first parent returns a UDP Miss message if
  all caches miss within two seconds.

10
Caching Resolution Protocol (contd)

• The cache will not wait for a home machine to
  time out.
• it will begin transmitting as soon as all of the
  parent and sibling caches have responded.
• The resolution protocols goal is for a cache to
  resolve an object through the source (cache or
  home) that can provide it most efficiently.
• This protocol is really a simple heuristic
• Fast response to a ping indicates low latency
• But bandwidth is more important for large objects.

11
Non-cacheable Objects/Security

• The wide variety of Internet information systems
  leads to a number of cases where objects should
  not be cached.
• Objects that are password protected are not
  cached. Rather, the cache acts as an application
  gateway and discards the retrieved object as soon
  as it has been delivered. ? can resolve security
  and privacy problems.
• CGI-Bins (server-side scripts)
• May limit the size of the largest cacheable
  object, so that a few large FTP objects do not
  purge ten thousand smaller objects from the
  cache.
• Caching subsystem does not prevent servers from
  encrypting or applying digital signature to their
  documents.

12
Cache Updates Problem

• Problems with caching
• Difficult to know if a cache object has been
  updated
• before its next use without checking (at least
  HEAD)
• No integrated mechanism in Web for remotely
  forced cache flush
• Can be controlled by object header (e.g.,
  Expires 0 Expires Thu, 16 May 2001 144030
  GMT).
• This mechanism only supports predictive
  expiration (says in advance how long a copy may
  be used).
• But what if unexpected change before expiration
  or unchanged persistence after that specified
  time?
• Cache Updates
• Based on Data access efficiency Use log of uses
  statistics (LRU Least Recently Used) triggered
  by a cache server.
• Cache consistency problem may occur when one or
  more cache servers maintain same copy of object.

13
The Economy of Cache Updates
14
Negative Caching

• To reduce the costs of repeated failures,
  negative caching is used.
• When a DNS lookup failure occurs, Harvest caches
  the negative result for five minutes (chosen
  because transient Internet conditions are
  typically resolved this quickly).
• When an object retrieval failure occurs, Harvest
  caches the negative result for a parameterized
  period of time, with a default of five minutes.

15
Replication Subsystem

• Motivations
• like to have(complete) regional copies with
  mechanism to ensure active consistency
  updates
• mirror-d (replication tool for Harvest using ftp
  mirror)

site2
site1
site3
Thin black mirror
Thick gray locally maintained master copies
16
Replication Subsystem (contd)

• Active consistent updates
• If a server changes its master copy, it notifies
  mirror sites.
• Harvest supports replication domains
• Mirroring within domain and carefully
  coordinated/synchronized
• Mirroring/replication between domains involves
  gradual propagation of changes (between sites
  responsible for inter-domain communication)

17
Replication Subsystem (contd)

• mirror-d replication tool weakly consistent
  replicated tree of files
• Motivation multiple copies for future access
• (e.g. Europe, North America) ?
  replication domain
• Problem maintaining data consistency
• Logical topology
• replication subgroups that coordinate consistency
  and internally share updates within subgroup
  domain.
• Physical issues(network bandwidth/usage)
• Help determine how replication domains propagate
  (flood) updates among its neighbors ? flooding
  handling issue (e.g., flooding in Peer to Peer
  network issues)

18
Replication Group
19
Replication Group (contd)

• Although Replication Domain members are stable,
  Pathways for inter-domain communication may
  change based on dynamic properties of server load
  and bandwidth

20
Physical vs. Logical Topology

• Logical inter-domain network topology is a subset
  of the
• full physical topology (and is dynamically
  re-configurable based on network load and
  bandwidth)

21
Replication in Broker World
22
Index Subsystem

• Please refer the papers for understanding
  indexing subsystem.
• GLIMPSE
• Uses Essence SOIF objects to create inverted
  index entries.
• NEBULA
• Supports hierarchical classification scheme ?
  automatic Yahoo classification.
• Provides views (pre-computed query responses) ?
  basically vector clusters w.r.t. a given query.

23
Harvest NG Architecture

• User agents goal directed
• extraction, analysis,
• even dialog
• Meta Brokers meta search
• collection/query fusion
• Brokers(Index, Search)
• Gatherers gathering and extracting (SOIF)
• Finders (Spiders) locate pages
• Content Provider (Web content pages)