Scalable ContentAddressable Networks

1
Scalable Content-Addressable Networks

• Prepared by
• Kuhan Paramsothy
• March 5, 2007

2
High-Level Overview

• Hash tables (map keys to values) are heavily used
  in building software applications
• The concept of a Content-Addressable Network
  (CAN) provides hash table-like functionality on
  Internet-like scales.
• CAN is
• Scalable
• Robust/Fault-tolerant
• Self-organizing
• Low-latency

3
Hash Tables and CAN

• A data structure that efficiently maps keys onto
  values
• CANs are a form of distributed, Internet-scale
  hash tables.

4
What CAN would do for us

• CAN would improve peer-to-peer systems
• Napster the process of locating a file is
  centralized
• Expensive to scale the central repository, single
  point of failure
• Gnutella decentralized the file location process
  (network self-organizes into an application layer
  mesh)
• Requests for files are done through flooding, not
  scalable, may not find content
• Conclusion P2P systems need a scalable indexing
  mechanism
• CAN would improve large data repositories
• These systems need efficient insertion and
  retrieval
• CAN would create large-scale name resolution
  services that dont use a naming scheme (ie. Not
  DNS)
• No more location-dependent naming schemes

5
Basic Operations Performed On CANs

• Basic Operations
• Insertion (of key,value pairs)
• Lookup (of key,value pairs)
• Deletion (of key,value pairs)
• Each CAN stores
• A piece (called a zone) of the entire hash table
• Holds information about a small number of
  adjacent zones in the table
• Routing in a CAN
• Done by intermediate CAN nodes towards the CAN
  node whose zone contains that key
• CAN Design is
• Distributed (requires no centralized control or
  coordination)
• Scalable (nodes hold only a small about of
  information that doesnt grow with the network)
• Fault-tolerant (nodes can route around failures)
• Doesnt require a naming hierarchy
• Is entirely Application Layer

6
CAN Design

• Centers around a virtual d-dimensional Cartesian
  coordinate space on a d-torus
• At any time, the entire coordinate space is
  dynamically partitioned among all the nodes in
  the system
• Each node owns a distinct zone

7
CAN Design (2)

• To store a pair, key K1 is mapped to P via a
  uniform hash function
• The pair is then stored at the node that owns the
  zone where P lies
• To retrieve an entry corresponding to K1, any
  node can apply the same hash function to map K1
  to P and get the corresponding value
• A node learns and maintains the IP addresses of
  those nodes that hold adjoining coordinate zones
• Efficient routing is critical to a useful CAN

8
Routing in a CAN

• Routing in a Content Addressable Networks works
  by following the straight line path through the
  Cartesian space from source to destination
  coordinates.
• A CAN node maintains a coordinate routing table
  that holds the IP address and virtual coordinate
  zone of each of its immediate neighbors in the
  coordinate space.
• Average Path Length (d/4)(n1/d)
• Individual Nodes Have 2d Neighbors
• Average Path Length Grows As O(n1/d)

9
Construction of a CAN Overlay

• The entire CAN space is divided amongst the nodes
  currently in the system
• Incremental construction process takes three
  steps
• The new node finds a node already in the CAN
• Using the CAN routing mechanisms, finds a node
  whose zone will be split
• The neighbors of the split zone must be notified
  so that routing can include the new node
• Bootstrapping There are CAN bootstrap nodes
  associated to a DNS domain name
• Node Insertion Affects Only O(number of
  dimensions) existing nodes

10
Maintenance of a CAN Overlay

• Node Graceful Departure node explicitly hands
  over its zone and the associated (key,value)
  database to one of its neighbors
• Node Abrupt Disappearance An immediate takeover
  algorithm ensures one of the failed nodes
  neighbors takes over the zone
• Under normal conditions, a node sends periodic
  update messages to each of its neighbors and a
  list of neighbors and their zone coordinates.
• Prolonged absence of an update message from a
  neighbor signals its failure

11
Design Improvements

• Basic CAN algorithm provides
• Low per-node state (O(d) for a d-dimensional
  space)
• Short path lengths (O(dn1/d) hops for d
  dimensions and n nodes)
• The problem is that there are application-layer
  hops, not IP-layer hops
• Latency of each hop might be substantial

12
Design Improvements (2)

• Improvement Multi-dimensioned Coordinate Spaces
• Increasing the dimensions of the CAN coordinate
  space reduces the routing path length and path
  latency for a small increase in the size of the
  coordinate routing table
• Path Length scales as O(d(n1/d))
• Fault-tolerance improves
• Improvement Multiple Coordinate Spaces (a.k.a.
  Multiple Realities)
• Maintain multiple independent coordinate spaces
  with each node in the system being assigned a
  different zone in the coordinate space (each
  coordinate space is a reality)
• Fault-tolerance improves
• Low per-node state (O(d) for a d-dimensional
  space)
• Short path lengths (O(dn1/d) hops for d
  dimensions and n nodes)
• Which is better?
• Increasing the dimensions

13
Design Improvements (3)

• Improvement Better CAN Routing Metrics
• Have each node measure the network-level
  round-trip-time RTT to each of its neighbors.
  Then route messages accordingly.
• Favors lower latency paths and avoids
  unnecessarily long hops
• Improvement Caching and Replication
• A CAN node can maintain a cache of the data keys
  it recently accessed
• A CAN node can replicate the data key at each of
  its neighboring nodes
• Both schemes need an associated time-to-live
  field, to eventually expire from the cache

14
Related Systems

• Domain Name System
• CANs are more general than the DNS because DNS
  closely ties the naming scheme to the manner in
  which a name is resolved to an IP address
• Peer-to-Peer
• A simple example is keys being analogous to a URL
• Will improve robustness
• Key difference is that content within the CAN can
  always be located by any other node because there
  is a clear home (point) in the CAN for that
  content and every other node knows what the home
  is how to reach it

15
Discussion

• Security?
• Better or worse with CAN?
• Any Other Design Improvement?
• Is The Communication Overhead Significant?

16
References

• A Scalable Content-Addressable Network,
  Ratnasamy, University of California Berkeley,
  http//www.sigcomm.org/sigcomm2001/p13-ratnasamy.p
  df