Using Gossip to Build Network Overlays.

18
Using Gossip to Build Network Overlays. Ken Birman Cornell University. CS5410 Fall 2008.

description

Using Gossip to Build Network Overlays. Ken Birman Cornell University. CS5410 Fall 2008. . Gossip and Network Overlays. A topic that has received a lot of recent attention Today we’ll look at three representative approaches - PowerPoint PPT Presentation

Transcript of Using Gossip to Build Network Overlays.

Page 1: Using Gossip to Build Network Overlays.

Using Gossip to Build Network Overlays.

Ken BirmanCornell University. CS5410 Fall 2008.

Page 2: Using Gossip to Build Network Overlays.

Gossip and Network OverlaysA topic that has received a lot of recent

attentionToday we’ll look at three representative

approachesScribe, a topic-based pub-sub system that runs

on the Pastry DHT (slides by Anne-Marie Kermarrec)

Sienna, a content-subscription overlay system (slides by Antonio Carzaniga)

T-Man, a general purpose system for building complex network overlays (slides by Ozalp Babaoglu)

Page 3: Using Gossip to Build Network Overlays.

ScribeResearch done by the Pastry team, at MSR

lab in Cambridge EnglandBasic idea is simple

Topic-based publish/subscribeUse topic as a key into a DHT

Subscriber registers with the “key owner” Publisher routes messages through the DHT owner

Optimization to share load If a subscriber is asked to forward a subscription, it

doesn’t do so and instead makes note of the subscription. Later, it will forward copies to its children

Page 4: Using Gossip to Build Network Overlays.

Architecture

20/12/2002 4

TCP/IPInternet

SCRIBEScalable communicationservice

Subscription managementEvent notification

PASTRYP2P location and routing layer

DHT

Page 5: Using Gossip to Build Network Overlays.

DesignConstruction of a multicast tree based on

the Pastry networkReverse path forwardingTree used to disseminate events

Use of Pastry route to create and join groups

20/12/2002 5

Page 6: Using Gossip to Build Network Overlays.

SCRIBE: Tree ManagementCreate: route to

groupIdJoin: route to

groupIdTree: union of

Pastry routes from members to the root.

Multicast: from the root down to the leaves

Low link stress Low delay

20/12/2002 6

groupId

join( groupId)

Multicast (groupId)

Root

join( groupId)Forwards two

copies

Page 7: Using Gossip to Build Network Overlays.

SCRIBE: Tree Management

20/12/2002 7

d13da3

65a1fc

d467c4: rootd471f1

Name space

26b20d

Proximity space

26b20d

65a1fcd13da3

d467c4: root

Page 8: Using Gossip to Build Network Overlays.

Concerns?Pastry tries to exploit locality but could these

links send a message from Ithaca… to Kenya… to Japan…

What if a relay node fails? Subscribers it serveswill be cut offThey refresh subscriptions, but unclear how

often this has to happen to ensure that the quality will be good

(Treat subscriptions as “leases” so that they evaporate if not refreshed… no need to unsubscribe…)

Page 9: Using Gossip to Build Network Overlays.

SCRIBE: Failure ManagementReactive fault toleranceTolerate root and nodes failureTree repair: local impact

Fault detection: heartbeat messagesLocal repair

20/12/2002 9

Page 10: Using Gossip to Build Network Overlays.

Scribe: performance1500 groups, 100,000 nodes, 1msg/groupLow delay penaltyGood partitioning and load balancing

Number of groups hosted per node : 2.4 (mean) 2 (median)

Reasonable link stress: Mean msg/link : 2.4 (0.7 for IP)Maximum link stress: 4*IP

20/12/2002 10

Page 11: Using Gossip to Build Network Overlays.

Topic distribution

20/12/2002 11

Gro

up S

ize

Topic Rank

Instant Messaging

Windows Update

Stock Alert

Page 12: Using Gossip to Build Network Overlays.

Concern about this data setSynthetic, may not be terribly realistic

In fact we know that subscription patterns are usually power-law distributions, so that’s reasonable

But unlikely that the explanation corresponds to a clean Zipf-like distribution of this nature (indeed, totally implausible)

Unfortunately, this sort of issue is common when evaluating very big systems using simulations

Alternative is to deploy and evaluate them in use… but only feasible if you own Google-scale resources!

Page 13: Using Gossip to Build Network Overlays.

Delay penalty

20/12/2002 13

0

250

500

750

1000

1250

1500

0 0.5 1 1.5 2 2.5 3 3.5 4 4.5

Delay Penalty Relative to IP

Cum

ulat

ive

Num

ber o

f Top

ics

Mean = 1.66Median =1.56

Page 14: Using Gossip to Build Network Overlays.

Node stress: 1500 topics

20/12/2002 14

Num

ber o

f nod

es

Total number of children table entries

Mean = 6.2Median =2

Page 15: Using Gossip to Build Network Overlays.

Scribe Link stress

20/12/2002 15

0

5000

10000

15000

20000

25000

30000

35000

40000

1 10 100 1000 10000

Num

ber o

f Lin

ks

Scribe

IP Multicast

Link stress

Maximum stress

Mean = 1.4Median = 0

Page 16: Using Gossip to Build Network Overlays.

AnycastSupports highly dynamic groupsSuitable for decentralized resource discovery

(can add predicate during DFS)Results (100k nodes/.5M network):

Join: 4.1 msgs (empty group); avg 3.5 msgs (2,500 members)

1,000 anycasts: 4.1 msg (empty group); avg 2.3 msgs (2,500 members)

Locality: For >90% of anycasts, <7% of member were closer than the receiver

20/12/2002 16

Page 17: Using Gossip to Build Network Overlays.

Fireflies

Fireflies.ppt

Page 18: Using Gossip to Build Network Overlays.

T-Man

T-Man