ChordNet: A Chord-based self-organizing super-peer network

Transcription

1 ChordNet: A Chord-based self-organizing super-peer network Dennis Schwerdel, Matthias Priebe, Paul Müller Dennis Schwerdel University of Kaiserslautern Department of Computer Science Integrated Communication Systems ICSY Outline Introduction Overlay structure Improvements

2 Introduction Peer to peer overlay network Based on Chord DHT Inherits basic structure Lots of improvements General purpose Not only for searching/locating content Also for transfering content 3 Chord Peer to peer overlay For storing key/value pairs (DHT) All peers arranged in a ring sorted by IDs Connections to predecessor and successort Shortcuts to nodes with ID + 2^n ( Fingers ) Routing to an ID uses farest finger preceeding the destination ID Keys are mapped to IDs using hash functions Nodes store all key/value pairs for Ids between them and their successor 4

3 Chord: Example 5 Chord Summary Efficient message delivery O(log n) Only unicast, no multicast, no broadcast Scales well with millions of nodes All nodes must be reachable Problems with Firewalls/NAT Does not consider latency 6

4 ChordNet Structure Separation: Super-Peers/Edge- Peers Super-Peer Reachable Inside the Chord ring Has list of its edge-peers Edge-Peer Behind Firewall/NAT, not reachable Connected to preceding and succeeding super-peer Communicates through its superpeers 7 ChordNet routing Only considering super-peers d Fingers in both directions Finger i to nearest node to (ID2^(i-1)) mod N Successor is Finger 1, Predecessor is Finger -1 Only slightly increases connection count Routing in both directions possible Routing can overshoot When fingers are optimal Average case: log(n)/3 hops (33% reduction) Worst case: log(n)/2 hops (50% reduction) 8

5 Multicast/Broadcast Multicast destination = ID list Broadcast destination = ID range Routing split list/range at each hop when needed Global broadcast has n-1 messages 9 Routing improvements Chord: Routing only based on IDs Routing will take expensive hops because of ID gain Better: Calculate routing gain (PRS) Latency can measured on connections Formula: gain = (old-dist new-dist) / latency-costs Even better: gain = (1/new-dist 1/old-dist) / latency-costs Chord: Decision is made by looking at own fingers Better: Look at finger tables of finger peers 2-Hop greedy routing (NoN) Additional messages needed for finger table exchange Both improvements can be combined 10

6 Routing improvement impact 11 Maintenance method Chord searches for fingers with routed requests Better: exchange finger-table with peers Needed for NoN rounting anyway Allows for quick start Combine predecessor and successor finger tables 12

7 Shortcut connections Problem: multiple hops are inefficient for big data transfers DHTs have no data transfers, new problem Traffic influences latencies, automatic load balancing with PRS Idea: establish shortcut connections Nodes detect high traffic routes using a trafficmatrix When HTR is detected, notification is sent to both peers Peers establish shortcut connection Shortcut connections are treated as fingers Result: HTR path length reduces automatically HTR peers are connected directly after a short while 13 Shortcut impact 14

8 ChordNet status Protocol specification Reference implementation in Java Able to run in simulation mode Addition features (work in progress) Publish/Subscribe API MessageQueue API Datagram API TCP-like API 15