Chord: A Scalable Peer-to-peer Lookup Service for Internet Applications

Chord Stoica et al., SIGCOMM '01 May 4, 2014 |
Emaad Ahmed Manzoor CS341: Advanced Topics in Data Management

Problem Motivation & metrics Solution The Chord Protocol Protocol Analysis
Provable guarantees Experimental Results Simulations May 4, 2014 | Emaad Ahmed Manzoor

May 4, 2014 | Emaad Ahmed Manzoor

May 4, 2014 | Emaad Ahmed Manzoor Map a key
to a node in the network containing the corresponding value

to a node in the network containing the corresponding value Keys are distributed across decentralized nodes

to a node in the network containing the corresponding value Nodes may arrive or leave at any time Keys are distributed across decentralized nodes

to a node in the network containing the corresponding value

to a node in the network containing the corresponding value Routing Tables

to a node in the network containing the corresponding value Communication Cost Routing Tables

to a node in the network containing the corresponding value Communication Cost Routing Tables Maintenance Cost

to a node in the network containing the corresponding value Communication Cost Routing Tables Scalability f(N) Maintenance Cost

May 4, 2014 | Emaad Ahmed Manzoor Keys are distributed
across decentralized nodes

across decentralized nodes Load Balance f (|k|/N)

across decentralized nodes Load Balance f (|k|/N) Asynchronous

May 4, 2014 | Emaad Ahmed Manzoor Nodes may arrive
or leave at any time

or leave at any time Availability

or leave at any time Availability + Asynchronous

or leave at any time Availability + Asynchronous = Correctness

May 4, 2014 | Emaad Ahmed Manzoor Scalability Load Balance
Availability Asynchronous Correctness Map a key to a node in the network containing the corresponding value Keys are distributed across decentralized nodes Nodes may arrive or leave at any time

Problem Motivation & metrics Solution The Chord Protocol Protocol Analysis
Provable guarantees Experimental Results Simulations May 4, 2014 | Emaad Ahmed Manzoor

May 4, 2014 | Emaad Ahmed Manzoor N N N
N N N Internet

May 4, 2014 | Emaad Ahmed Manzoor N N N
N N N Internet Publisher key = k value = v

May 4, 2014 | Emaad Ahmed Manzoor Publisher key =
k value = v N N N N N N Internet Client key = k value = ?

k value = v N N N N N N Internet Client key = k value = ? Assignment Load Balancing Asynchronous

May 4, 2014 | Emaad Ahmed Manzoor Consistent Hashing Distribute
keys across nodes

May 4, 2014 | Emaad Ahmed Manzoor Consistent Hashing key_identifier
= hash(key) = 110001 m bits, m = 6

= hash(key) = 110001 node_identifier = hash(node IP) = 100101 m bits, m = 6

= hash(key) = 110001 node_identifier = hash(node IP) = 100101 m bits, m = 6 Identifier ring identifier % 2m = 0 to 63

= hash(key) = 110001 node_identifier = hash(node IP) = 100101 m bits, m = 6 Node for k successor(key_identifer) Identifier ring identifier % 2m = 0 to 63

May 4, 2014 | Emaad Ahmed Manzoor Load Balancing with
Consistent Hashing N nodes, K keys, with high probability: - Each node has at most (1 + e)K/N keys - On arrival or departure of a node, O(K/N) keys exchange hands

Consistent Hashing (a) N = 10,000 (b) k = 500,000

Consistent Hashing (a) N = 10,000 (b) k = 500,000 Large variation in the number of keys per node

Consistent Hashing (a) N = 10,000 (b) k = 500,000 Many nodes have no keys

May 4, 2014 | Emaad Ahmed Manzoor Reason for The
Load Imbalance

Load Imbalance Node identifiers do not uniformly cover the identifier space

Load Imbalance Node identifiers do not uniformly cover the identifier space Solution Virtual Nodes

May 4, 2014 | Emaad Ahmed Manzoor Load Balance with
Virtual Nodes

Virtual Nodes Load balance improves with more virtual nodes per real node

Virtual Nodes Load balance improves with more virtual nodes per real node Cost of routing information per node also increases

k value = v N N N N N N Internet Client key = k value = ? Assignment Load Balancing Asynchronous

k value = v N N N N N N Internet Client key = k value = ? Lookup Scalability Asynchronous Assignment Load Balancing Asynchronous

May 4, 2014 | Emaad Ahmed Manzoor Routing information: O(1)
Communication: O(N)

May 4, 2014 | Emaad Ahmed Manzoor m entries

May 4, 2014 | Emaad Ahmed Manzoor Routing information: log(N)

May 4, 2014 | Emaad Ahmed Manzoor Communication cost: ?

May 4, 2014 | Emaad Ahmed Manzoor Each node can
forward a request at most halfway around the identifier circle

May 4, 2014 | Emaad Ahmed Manzoor Expected number of
query steps if node n wants to find the predecessor p of key k = ?

query steps if node n wants to find the predecessor p of key k = ? 0000 1000 1011

query steps if node n wants to find the predecessor p of key k = ? 0000 1000 1011 d 0 = 1011

query steps if node n wants to find the predecessor p of key k = ? 0000 1000 1011 d 1 = 0011

query steps if node n wants to find the predecessor p of key k = 0.5 x log(N)

May 4, 2014 | Emaad Ahmed Manzoor Routing information: O(log(N))
Communication cost: O(log(N))

May 4, 2014 | Emaad Ahmed Manzoor Lookup Cost in
Practice 212 node network

Practice Path length = query steps. Grows as log(N) 212 node network

Practice Path length = query steps. Grows as log(N) Mean path length = 0.5 x log(N) 212 node network

Practice Path length = query steps. Grows as log(N) Mean path length = 0.5 x log(N) High variance in path length?

k value = v N N N N N N Internet Lookup Scalability Asynchronous Assignment Load Balancing Asynchronous Client key = k value = ?

k value = v N N N N N N Internet Lookup Scalability Asynchronous Assignment Load Balancing Asynchronous Client key = k value = ? Joins/Failure Availability Correctness Scalability

May 4, 2014 | Emaad Ahmed Manzoor Stabilize protocol -
Stabilize Updates successors - Notify Updates predecessors - Fix fingers Updates finger tables

May 4, 2014 | Emaad Ahmed Manzoor What if a
lookup happens now?

May 4, 2014 | Emaad Ahmed Manzoor Correctness of Lookups
with Node Joins 3 cases - Finger table entries are current, successors are correct

with Node Joins 3 cases - Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result

with Node Joins 3 cases - Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result - Finger table entries are stale, successors are correct

May 4, 2014 | Emaad Ahmed Manzoor 3 cases -
Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result - Finger table entries are stale, O(N) lookup cost successors are correct Correctness of Lookups with Node Joins

May 4, 2014 | Emaad Ahmed Manzoor 3 cases -
Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result - Finger table entries are stale, O(N) lookup cost successors are correct Correct lookup result Correctness of Lookups with Node Joins

with Node Joins 3 cases - Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result - Finger table entries are stale, O(N) lookup cost successors are correct Correct lookup result - Successors are incorrect

with Node Joins 3 cases - Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result - Finger table entries are stale, O(N) lookup cost successors are correct Correct lookup result - Successors are incorrect Lookup fails

with Node Joins 3 cases - Finger table entries are current, O(log(N)) lookup cost successors are correct Correct lookup result - Finger table entries are stale, O(N) lookup cost successors are correct Correct lookup result - Successors are incorrect Lookup fails Application layer pauses and retries failed lookups

May 4, 2014 | Emaad Ahmed Manzoor Performance of Lookups
with Node Joins

with Node Joins With high probability, O (log(N)) new nodes will land between any two existing nodes

with Node Joins With high probability, O (log(N)) new nodes will land between any two existing nodes Lookups remain O(log(N))

May 4, 2014 | Emaad Ahmed Manzoor Node Failure

May 4, 2014 | Emaad Ahmed Manzoor Node Failure Each
node maintains a successor list of size r

with Node Failure Mean path length with successor lists = 0.5 x log(N) - 0.5 x log(r) + 1

with Node Failure

with Node Failure Mean path length with successor lists = 0.5 x log(N) - 0.5 x log(r) + 1 Eliminates the last 0.5 x log(r) hops on average

with Node Failure Mean path length with successor lists = 0.5 x log(N) - 0.5 x log(r) + 1 Access the predecessor of the key k from the node n' retrieved from the successor list

May 4, 2014 | Emaad Ahmed Manzoor Scalability Load Balance
Availability Asynchronous Correctness Map a key to a node in the network containing the corresponding value Keys are distributed across decentralized nodes Nodes may arrive or leave at any time Chord

May 4, 2014 | Emaad Ahmed Manzoor .

Chord: A Scalable Peer-to-peer Lookup Service f...

Chord: A Scalable Peer-to-peer Lookup Service for Internet Applications

Transcript

More Decks by Emaad Manzoor

Other Decks in Science

Featured