Pilot: Probabilistic Lightweight Group Communication System for Ad Hoc NetworksAuthored by Luo, Eugster, and Hubaux

Presented by Jin-Hee Cho

Group Communication System (GCS) in Ad Hoc Networks GCS: A useful infrastructure on which various

reliable distributed computing functions can be built.

Mobility management Distributed management of cryptographic

keys or certificates Access control Key management

Overview of Paper Pilot provides probabilistic reliability for: Multicast: RDG & RRDG(R2DG) Use gossip mechanisms Data Sharing: PAN Use probabilistic quorum systems Present analytical results & simulation results

using ns-2 Show a tradeoff between reliability (Rd) and

efficiency (Nl)

Related Work: Gossip-Based Probabilistic Reliable Multicast

Probabilistic reliable multicast protocols Reduce the protocol overhead by sacrificing safety

guarantees by using gossip-based dissemination scheme.

Examples Probabilistic Broadcast Lightweight Probabilistic Broadcast Anonymous Gossip (AG)

Related Work: Probabilistic Quorum Systems State Machine approach “write all- read one.” Synchronizing a set of replica to handle all updates and queries

in the same way. Perfect guarantee in theory BUT too expensive in reality. Original (Strict) Quorum System “write many-read many.” A subset of quorums, each consisting of a subset of server

replica, i.e. read quorum and write quorum. Ensure the intersection of read and write quorums. Probabilistic Quorum System Relax the intersection property. i.e. no intersection may happen. First introduced by randomized database groups [Hass99].

Related Work: Data Management in Ad Hoc Networks 7DS [Papadopouli01] vs. Pilot Diffusion scheme under different network

environment: low density vs. high density Data dissemination: node mobility vs. a

gossip-based protocol. [Hara01] & [Wang02] vs. Pilot Mobility Model: network partition/prediction

vs. random way point model.

Model

An ad hoc network with a set N of nodes Every node with unique id. Node failure/recovery Underlying unicast protocol: DSR Address two fundamental problems: multicast

and data sharing.

i N

Metrics for Reliable Multicast Reliable Multicast Protocol: Disseminates packets within a multicast group, G N Reliability Degree of Single Packet Dissemination

(Rds): the fraction of group members that receive the packet sent by a certain member.

Reliability Degree of Continuous Dissemination (Rdc): the fraction of all packets that are received by a certain member with the rate λ0

Both metrics are described by respective cdf F(x): [0,1][1,0] the probability that Rds or Rds is at most x.

Metric for Reliable Data Sharing Reliable Data Sharing Service Storage set, STS N where is a set of nodes. ρ is a set of access protocols for STS. STS holds shard data in a replicated fashion using the consistency

model of data replication called “shared private.” “shared private”: the object is owned by a particular node. Only this

node can modify the object while others can read it. Given access rates λu and λq for updates and queries, the data

sharing service is probabilistically reliable in nature if a query access ρq (STS, λq) obtains, with a certain probability, the latest version of a data object resulting from an update access ρu (STS, λu).

Reliability Degree of Access (Rda): Probability that a query operation acquires the most recent update of the corresponding data object, considering both node and channel failures.

Metric for Overhead & Goal Network load (Nl) Average number of unicast packet*hop per

multicast packet to achieve a certain Rds or per unit time to achieve a certain Rdc or Rda

Goal: design a set of protocols that achieve a high reliability degree Rd (Rds, Rdc, & Rda) even under high λ0 while incurring reasonable overhead, Nl.

Show an efficient tradeoff between Rd and Nl.

Layered Architecture of Pilot Route Driven Gossip (RDG):

a gossip-based probabilistic multicast protocol

Reliable RDG (R2DG): devised for continuous packet dissemination and detects packet loss

Probabilistic quorum system in Ad hoc Network (PAN): Any node in STS is termed server while the rest of the nodes are termed clients of STS.

Data query and updates: arbitrary servers in STS Message dissemination within STS: RDG.

Gossip-based multicasting in RDG

Message exchange for updating and querying the STS in PAN

RDG: Basic Pilot Multicast Protocol Pure gossip scheme basis Use available routing information Random subview works well Each packet with pid [gid, sid, seq] Four data structures Data management: pidList, Buffer Membership management: gidList, Views AView (active view), PView (passive view), RView

(remove view) Each node has four subrecords (pidListi

gid,Bufferigid,

gidListigid, Viewi

gid)

RDG Operations: Join session GROUPREQUEST message Update AView by all other members GROUPREPLY with probability Preply

Update AView by initiator Maintain AView and Pview updated by

recording the route of each incoming packet Reinitiate if the size of AView drops under

threshold.

RDG Operations: Gossip/Leave session Three protocol parameters in GOSSIP task F: fanout is the number of gossip destinations randomly

selected from the AView for each gossip emission. τa : the quiescence threshold means that each packet will

be removed from Buffer after having been gossiped for τq

rounds by individual nodes. Thus, it limits the number of gossip rounds.

τq : the age threshold limits the propagation range of each packet. Thus, it indicates how many times a packet is repeatedly relayed by a certain group member.

RDG Operations: Gossip/Leave session

Multicast

Node Leave

RDG Operations: Gossip/Leave session

gossiping gossiping

Update views

Packet Emission

RDG Operations: Gossip/Leave session Packet Reception

R2DG: Continuous Packet Multicasting Service

Same data structure as RDG except larger Buffer size to detect missing packets

Detects missing packets by examining the pid of sequence of received packets

Pull packets or piggybacks the pull information.

Multicast & Pull task

R2DG: Continuous Packet Multicasting Service

Packet reception and the response to pull

PAN: Reliable Data Sharing Service Client protocol: requests to an arbitrary server in the

STS, which is terms an agent. One-to-one connection. Each message: mid [sid, oid, ver] Server protocol: maintains a quorum system building

upon the STS with the support from the underlying RDG protocol.

Nominal quorum size: Real quorum size: Read quorum: R Write quorum: W

?ˆ?

PAN: Server Update Protocol

UPDATE emission

UPDATE reception

PAN: Server Query Protocol

QUERY emission (Agent Servers)

PAN: Server Query Protocol

QUERY reception (Servers Agent)

PAN: Server Query Protocol

REPLY reception at an agent

Examples of Protocol Operations

15 9, 1015 5, 3910, 1310 2, 8

9 3, 551, 19101, 131310, 1521, 8

All members receive the message.

F = 2

τa= 2

|G| = 10

N = 20

Examples of Update/Query in PAN

|STS| = 25

N = 50

M25: Update

M27: Query

M12: the intersection of write and read quorum

Analysis Model |G| = n members/servers Gossiping operations: a uniform random selection of F members out

of n. Infected member: A member that has received a certain packet. Susceptible member: A member that has NOT received a certain

packet. Infectious member: A infected member who keeps gossiping the

packet. A node gossips in synchronous rounds (Tms)

pf: failure probability for each hop along a routing path

(pfc « pfmo ≈ pf) failure due to network condition. H: the number of hops pe: server unavailability due to failure at any instant time.

Analysis Model

No consideration for R2DG and client protocol in PAN. Overall access rate λ0 = λq + λu

The dissemination process of the server update by RDG Consider only the second query to a data object that was

modified by the most recent update, while considering the first query as happening before the update.

Stochastic Behavior of RDG Sr: the number of members infected with the packet after

round r.

1 0

1

{0,..., } Pr{ 0} 1 0

[ , ,..., ]

:

{0,..., } ... {0,..., }

q

q

r r

Tr r r r r

S n S for r

S S S

state space

n n

S

Stochastic Behavior of RDG: Recurrence Relation (1)

1 1

( )

( | , )

( )(1 )

0

r r r r

n i k j i k n jj i

P S j S i S S k

q q j i

j i

n

k = j-i i (# of infectious members)

(# of infected members)

the probability of noninfection

the probability that a certain susceptiblle member

is not infected by any of the k infectio

qr r

r

k

S S k

S i

q

q

us members.

Group

r+1 round

r round

The probability of having successfully infected members in round r+1 among susceptible members in round r by infectious members of round r.

Stochastic Behavior of RDG: Recurrence Relation (2) & (3)

The probability of having i infected members in round r.

Stochastic Behavior of RDG: Computation of p The probability of infection p can be estimated under two conditions: (i) the considered node is chosen as the gossip destination and (ii) the gossip message is successfully received.

Stochastic Behavior of RDG: Reliability Degree Rds and Rdc

p1 is the probability that a certain member is infected by receiving a single packet in round r.

M: the number of packets in a stream.

Stochastic Behavior of RDG: Network Load (Nl)

Network Load for a single packet dissemination

Network Load for continuous packet dissemination

Stochastic Behavior of PAN: Reliability Degree Rda

pdf of real read quorum

pdf of real write quorum

probability that a query happens within a certain period (between round r and r+1)

The probability for read quorums to intersect with write quorums in each round r

Stochastic Behavior of PAN: Network Load (Nl)

Simulations: Model and Parameters ns-2 with the Monarch Project wireless and mobile extension. Simulation area: 1km2

Random waypoint model Gossip period: 200ms λ0 = 5 pkt/s λ0 = 8 λu

Vary F and Ξr pe = 0.01 Simulated time: 400 seconds N = 100, |G| = n = 50 Paired speed/pause time: 2m/s, 5m/s, 10m/s and 20m/s, and

10s, 20s, 40s, and 80s.

Single Packet Dissemination Reliability Rds

Speedmax = 2m/s Timepause = 40s (a) vary F with τa= 1 (2) vary τa with F = 3

Continuous Packet Dissemination Reliability (Rds) and Network Load (Nl)

|G| = 50. mobility is varied. R2DG-pull mechanism.

Continuous Packet Dissemination Reliability (Rds) and Network Load (Nl)

Speedmax = 1m/s, Timepause = 10s Scalability Effect

Impact of λ0 on PAN Performance

F = 2, ξR = 4, relatively stable during 1.5/s< λ0 <3/s

Access Reliability (Rda) and Network Load (Nl)

Normal density network, F = 2, ξR = 4

Conclusion

To realize probabilistic reliable group communication in mobile ad hoc networks, two fundamental problems are studied, namely multicast and data sharing.

Pilot uses gossip mechanism and probalistic quorum systems.

Analytical results and simulation results.

Questions?