Chord

1
Chordstructured overlay network for non-uniform
load-distribution

• THORSTEN SCHUTT,
• FLORIAN SCHINTKE,
• ALEXANDER REINEFELD

2
Introduction

• We present a lookup method that does the routing
  without DHTs. It has the same favorable
  properties and performance as Chord but allows
  more expressive queries.

3
Lookup services

• We distinguish three kinds of lookup services.
• Name servicesprovide a mapping between names and
  attributes.
• Directory services Given a list of attributes, a
  directory search returns name or address of that
  item.
• Discovery servicesAllow services to register
  themselves in spontaneous networks for later
  lookup.

4
General layout of a lookup services
5
Performance Metrics

• We distinguish four kinds of performance metrics.
• Routing performance the average number of
  network hops traversed for finding objects.
• Entry load the number of key-value pairs each
  node is responsible for.
• Query load the average number of queries that
  are processed by a node in a given time.
• Routing load amount of routing operations
  performed by a node in a given time interval.

6
Traditional Chord
7
Chord

• The key idea of Chord is to substitute Chords
  hash function by a key-order preserving function
  and adjusting the finger tables accordingly.

8
Chord
9
Handling non-uniform distributions

• Chord is based on consistent hashing the
  following theorem holds true.

10
Handling non-uniform distributions

• density function

11
Handling non-uniform distributions

• To describe the node distribution we define the
  density function d(x) over the key space. It
  gives for each point x in the key space the
  reciprocal of the width of the corresponding
  interval.

12
Handling non-uniform distributions
13
Pointer Placement Algorithm
14
Pointer Placement Algorithm
15
Pointer Placement Algorithm
16
Pointer Placement Algorithm
17
Pointer Placement Algorithm
So, to calculate the 2nd (ith) pointer, a node
simply asks his neighboring node (i-1th pointer)
for his 1st (i- 1th ) pointer. In general, the
pointer of one level is set to the neighbors
neighbor on the next lower level. On the lowest
level, the pointer references to the direct
neighbor.
18
Empirical results

• For the simulation we used a corpus of 629,228
  English words 7. Apart from words, that are
  used in daily life, the corpus also contains
  short sequences of symbols/digits.

19
Routing Performance.
20
Adaptation Speed
21
Adaptation Speed
22
Adaptation Speed
23
Conclusion

• We presented an enhancement to the Chord
  algorithm that performs a data lookup without
  using DHTs.
• Moreover, they provide the means for handling
  more complex queries, like relational or
  attributed queries.