Object Naming

1
Object Naming Content based Object Search

• 2/3/2003

2
System Architecture

• Due to requirements for scalability, reliability
  and performance, hybrid architecture is used for
  the system.
• Regional manager provides centralized management
  of OSD devices, clients and objects within the
  region.
• P2P relationship is maintained among regional
  managers. Object location among regions is based
  on Distributed Hashing Table (DHT). DHT has good
  scalability, and is used by Tapestry, Pastry,
  CAN, Chord
• Current DHT-based systems do not consider
  locality, peers various processing power, and
  network topologies, including bandwidth, LAN and
  WAN, enterprise special requirement.

3
(No Transcript)
4
Object Placement

• Object is organized within region for efficient
  creation, update, and search.
• Object is stored within local region, and object
  or its metadata is also stored out of region.
• Dynamic replication/migration is a major
  component of object placement.

5
Object Naming

• In our proposal, object is identified with Globe
  Unique Identifier (GUID, such as
  f81d4fae-7dec-11d0-a765-00a0c91e6bf6 )
• Clients, including users and traditional file
  systems, may need symbolic object name. Same
  object may have different names in the light of
  different users. Same object name may refer to
  different objects in different environments.

6
Object Naming (Cont.)

• Mapping between GUID object name is a must.
  This mapping should be distributed, scalable,
  efficient, reliable and secure.
• Mapping also supports clients individual view of
  objects the client can access. GUID is in flat
  name space, but clients view of objects may have
  hierarchy-tree structure.
• Mapping between clients accessed object name and
  GUID will be cached for performance.

7
Object Retrieval based on Partial Information

• Client may want to access an object, but only
  knows about partial name or some key words of
  object contents. DHT-based P2P systems have good
  scalability with poor query facility. Currently,
  these systems only support exact match.

8
Object Retrieval based on Partial Information
(Cont.)

• Options
• Centralized index management
• Flooding search
• GUID resolution based on DNS like mechanism
• Content based routing on overlay networks
• Flooding search is not the solution. Gnutella
  does not scale well.
• DNS has many known problems, including single
  point of failure.

9
P2P Content Based Routing

• Goal based on key words, object request is
  efficiently routed to regional manager that holds
  the object.
• Number of regional mangers contacted should be as
  few as possible.
• With object replicas distributed among
  regions/OSD devices, nearby copy should be found.
  Definition of nearby copy considers bandwidth,
  object size, node processing power as well as
  node state.
• Role based Access Control should be honored.
• Object should be found as long as it exists in
  the system.

10
P2P Content Based Routing (Cont.)

• Efficient indexing on key words of objects need
  to be constructed, distributed and its freshness
  will be maintained. One useful index scheme may
  rely on a unique hashing function.
• Object request can be routed based on the
  constructed indexing.
• Approximate routing based on statistics is a
  possible solution.

11
(No Transcript)
12
P2P DB Systems
P2P
DB
Flexibility ? ?
Decentralized ? ?
Strong Semantics ? ?
Powerful query facilities ? ?
Fault Tolerance ? ?
Lightweight ? ?
Transactions Concurrency Control ? ?
13
Focusing Issues

• Efficient object placement mechanism within a
  region and among regions. It could be an
  innovative DHT, or indexing scheme.
• Scalable object search based on partial match,
  minimizing bandwidth , processing costs.
• Query processing optimization, including caching,
  client satisfaction.