Chord

Ólafur Helgason

VP of Engineering @ OZ

Papers We Love

Reykjavik University

19 nov 2014

Why Chord?

• Super-cool and simple
• Theory & practice
• Influential paper & topic
• Personal reasons - nostalgy

Presentation overview

• Background
• Distributed Hash Tables
• Chord
• Lookup and churn in Chord
• Applications
• Discussions - Yes you!

2001

Interest in Distributed Systems

Distributed Hash Table

• Same interface as normal hash table
  • keys map to values
• Keys are assigned to nodes
• Node stores all values for its set of keys
  • Hash table buckets are nodes in a netw
• Chord one of the first implementations
  • CAN, Tapestry, Pastry, Kademlia

Wtf do I care?

• Benefits
  • Decentralized
  • Scalability
  • Fault tolerance
  • Reliability
• Drawbacks/challenges
  • Nodes leave/fail
  • Nodes join

Chord: One ring to rule them all

lookup(key) -> node

One ring to rule them all

• Each node responsible for 1/n of keyspace
• Half of neighbours keyspace on join
• Split keyspace on leave
• Simple lookup
  • Each node queries neighbour
  • Node state (successor) ~ O(1) 
  • Lookup ~ O(N)

Example: m=6, 10 node, 5 keys

Simple lookup (non optimal)

Improve lookup performance

• Finger table
  • At most m entries ~ log(N)
    • keyspace [0, 2^(m-1))
  • Entry i
    • successor(n + 2^(i-1))

Finger table

Improved lookup

Churn

• Join
  • new node (N26) find its successor
  • receives keys from successor
  • each node periodically updates finger table
• Leave
  • Transfer keys to successor & notify predecessor

Overview

• Simple: lookup(key) -> node
• Scalable
  • State per node ~ O(log(N))
  • lookup performance ~ O(log(N))
• Provable correctness
  • Even under churn

DHT Applications/implementations

• BitTorrent distributed tracker 
• Overlay multicast
• Corel CDN
• Amazon Dynamo
  • Cassandra, Riak

Amazon Dynamo

• Each virtual node gets random ring position
• Data assigned to vnode such that
  • vnode = successor(key)
• Data replicated to N-1 next successors
  • N is number of replicas
  • Coordinator to 'tune' replication across different physical nodes 
• Fault-tolerance
  • Nodes can fail
  • Can read and write under network partitioning
    • Write not propogated to all nodes - inconsistency!