BATON Papers Overview

1
BATON Papers Overview

• Jagadish, Ooi, and Vu BATON A Balanced Tree
  Struture for P2P Networks, VLDB 2005.
• Jagadish et al., Speeding up Search in P2P
  Networks with a Multi-way Tree Structure, SIGMOD
  2006
• Presented by Nick Taylor, October 4, 2006

2
Overlay Networks

• Add additional routing protocol on top of network
  routing (i.e. TCP/IP)
• Need to be resilient to node failures,
  intermittent connectivity
• Redundancy in data and metadata
• Most popular is Distributed Hash Table (DHT)
• Pastry
• Chord

3
Chord Overview

• Randomly assign each peer an ID between 0 and 2m
• Hash key to find ID for an item
• Next peer after ID (mod 2m)is responsible for
  storing item
• Each peer has routing table to subsequent peers
• Table is used for incremental routing

4
Chord Routing Example (m 6)
Figures poached from Stoica et al., Chord A
Scalable Peer-to-peer Lookup Service for Internet
Applications, SIGCOMM 2001.
5
DHTs The Good, the Bad and the Ugly

• Good
• Efficient routing algorithms
• Easy to understand
• Simple replication methods
• Bad
• Load balancing depends on good choice of hash
  function
• Ugly
• Can only do exact ID lookups, since items are
  clustered in key space
• Would like to do range and prefix searches

6
BAlanced Tree Overlay Network
Note that routing table for a node m contains
details about nodes m-22, m-21, m-20, m20, m21,
sort of like Chord.
7
Lookup Algorithm
Lookup for 74 starting at node h
8
Lookup Algorithm (2)

• Rarely moves up tree
• If at leaf node
• If value is higher up tree
• At most log N messages sent to do a lookup
• Range query finds start of range using lookup
  procedure, then follows links between nodes
• Insertion and deletion use lookup procedure

9
Node addition and removal rules

• Need to preserve balanced structure of tree
• Addition forward request to node with full
  routing table but at most one child
• Removal find leaf node that can be moved without
  causing imbalance

10
Load Balancing
a

• May need to redistribute ranges between nodes
• Simple case just affects immediate neighbors
• General algorithm
• Identify overloaded node g
• Identify lightly loaded node f
• Transfer fs data to c
• Add f as child of g
• If tree is unbalanced, perform rotation
• Need to update routing table
• Still cheaper than moving data around

c
b
f
d
e
g
e
c
d
a
g
b
f
11
Experimental Results

• Maintenance slightly better than Chord
• Performance slightly worse than Chord
• Root node does not get overloaded
• Leaves get slightly more traffic than root

12
BATON

• Generalizes techniques from binary trees to m-ary
  trees
• Improves query performance
• Increases maintenance cost

13
Fault tolerance and load balancing

• m-ary trees are harder to partition
• m-ary trees have more leaf nodes
• Easier to do local load balancing

14
Conclusions

• DHTs and BATON each have their strengths
• DHTs simplicity, slightly better performance,
  easy replication
• BATON lower maintenance overhead, range queries
• m-ary trees are often better than binary trees
• Depends on frequencies of queries and updates