1. A Scalable Content Addressable Network
(Sigcomm 2001)
Authors: Sylvia Ratsanamy, Mark Handley, Paul Francis
Richard Karp, Scott Shenker
Ehsan Foroughi

2. Outline Introduction Design Improvements Logical Environment
Find and Insert Operations
Departure and Recovery
Improvements

3. Introduction Large Scale Addressing/Storage Management Key Features:
Distributable/Scalable
Fault Tolerance
Fast Access vs. Low Overhead
Hash Tables: How to distribute them?

4. Basic Design Logical Hash Space
Keys hash to a d-dimensional Cartesian co-ordinate space on a d-torus
The space is partitioned into distinct zones
Zones are d-dimensional rectangles
Zones are neighbor if share in (d-1) dimention

5. d-dimensional Torus Y X
1.0
0, 0
0.5

6. Routing
Queries are hashed to the space, then routed to the node that handles the destination point on the virtual space
Each node maintains information about its neighbors
Routing is greedy, i.e. the query is forwarded to the neighbor with least distance to destination

7. Example of Routing

8. Insertion Find an arbitrary node of the system
Pick a random key and hash it to some point
Find the owner of the zone using the routing mechanism
Start to Join
Split P

9. Insertion cont’d Split the destination zone
Take over a half of the new zone
Update the neighbor information lists both on new zones and all neighbors of the old zone
New Zone

10. Operation Cost Average path length is (d/2)(n1/d) hops
Average neighbor list is 2d
Letting d=log n both average path length and neighbor list will be O( log n )

11. Departure & Recovery Departure can be
With previous warning
Sudden death
In case of normal departure, the node will explicitly hand out the zone to one of its neighbors and the depart
Under normal conditions, nodes transmit update message so that neighbors will know when someone is dead using a timeout timer and ping!

12. Recovery
In case of abnormal departure, the neighbors decide who should TAKEOVER the zone (node with smallest zone)
In both case, some nodes may handle more than one zone temporarily
Massive cluster failures can bring CAN to an inconsistent state! (not very probable)

13. Zone Reassignment Fragmentation problem due to departure of nodes
A background algorithm reassigns zone and tries to merge-up zones to reduce the number of zones each node handles

14. Improvements Multi-Dimension Metrics Multi-Coordinate Spaces
Overloading the Zones
Multiple Hash Functions
Topologically Sensitive Construction
Uniform Partitioning
Caching

15. Conclusion Drawbacks Applications Not possible to increase dimension
Physical neighboring vs. logical neighboring
Applications
P2P File Sharing
DNS …