PeerDB: A P2P-based System for Distributed Data Sharing

1
PeerDB A P2P-based System for Distributed Data
Sharing

• Wee Siong Ng, Beng Chin Ooi, Kian-Lee Tan, Aoying
  Zhou

Shawn Jeffery CS294-4 Peer-to-Peer
Systems 11/05/03
2
Overview

• A P2P database system
• Allows content-based search
• No global schema
• Utilizes mobile agents
• Provides flexibility and extensibility
• Dynamically adjusts topology

3
Background P2P vs Distributed Databases
P2P DDBMS
Membership Ad-hoc Controlled
Schema No global schema Shared (or at least some way to mediate)
Query result set Incomplete Complete
Content location Word of mouth Shared catalog
4
BestPeer

• Generic P2P platform
• Mobile Agents
• Carry code and data
• Collect stats
• Security issues?
• Dynamic Reconfiguration
• How does this compare to Gia?
• Location Independent Global Names Lookup (LIGLO)
  Servers
• Small number
• Provides a global identity for peers and peer
  status
• Why not use a DHT/KBR/DOLR?

5
BestPeer Security

• Private and sharable data
• Agents only able to access sharable data
• Does this adequately restrict the power of mobile
  agents?
• Communications on the wire also encrypted
• Whats missing?

6
Architecture
Sharable Data
Local Data
Database
7
Schema Mediation

• Problems with supporting SQL queries
• No global schema information
• Different nodes could name the same
  table/attribute differently (len, length)
• Solution User supplies metadata for each
  relation name and attribute
• Users expected to do a lot
• Formula based on matching relation keywords and
  attribute keywords to determine if a query
  matches a table
• What about other schema mediation work (such as
  Piazza)?

8
Local Query Processing Phase I

• Master Agent coordinates the entire affair
• Check Local Dictionary for matching relations
• Use the relation matching strategy even for the
  local DB
• Create Relation Matching Agents and flood to
  all neighbors
• Wait for responses
• Display results to user as they arrive

9
Local Query Processing Phase II

• User selects the relations he/she wants
• Create a Data Retrieval Agent
• Rewrite query in terms of new relations
• If local, submit SQL to local db
• Contact remote nodes directly to access the data
• Creates remote join plans locally - optimization?

10
Remote Query Processing

• Phase I Find relations
• Relation Matching Agents flood with TTL
• Check Export Dictionary for a match
• Return matches directly
• Phase II Get data
• Data Retrieval Agent submits SQL to DBMS
• Return data to the requesting node directly
• Run further data processing before returning
• Again, security issues

11
Statistics

• Master Agents monitor stats in the network
• Keywords for some relations returned during Phase
  I
• Update metadata
• Number of objects returned for selected relations
• Can be used for topology change decisions
• Use most recently returned results as metric to
  determine who to connect with
• Frequent updates might need to change neighbors
  after each result returned

12
Caches

• Cache all query results locally
• Soft state
• LRU replacement
• Users choose which copy they want
• Only provided with peer id and an indication of
  which is the source
• What about timestamp, etc?
• Again, user heavily involved

13
Relation Matching Performance

• Significant tradeoff between precision and recall
• Which is more important?
• Is their approach acceptable?

14
Experimental Methodology

• Compare P2P Model vs Client/Server model
• CS returns via the search path (?)
• Compare static vs reconfigurable networks
• Compare agent vs message based approach
• 32 Nodes
• Is this enough?

15
Evaluation Scenarios (Metrics?)

• Fixed set of nodes
• Easily test P2P protocols, Reconfiguration
  strategies
• Latency
• Quality and Quantity
• What else is important?

16
Performance

• As you increase the amount of storage on each
  node, latency decrease
• Due to caching
• In general, reconfiguration performs better
• Response times O(1 Minute)
• Is this acceptable?
• Agent based shown to be better
• What if agent produces more data than it
  processes?

17
Discussion A P2P DBMS?

• PeerDB represents a tiny step towards a P2P DB
  (also PIER, Piazza)
• What does it do right?
• What else is needed?
• Is it ideal to have a P2P DB?
• Is it feasible?