1 eventual leader election in evolving mobile networks luciana arantes 1, fabiola greve 2,...
TRANSCRIPT
1
Eventual Leader Electionin Evolving Mobile Networks
Luciana Arantes1, Fabiola Greve2, Véronique Simon1, and Pierre Sens1
1Université de Paris 6 (LIP6) / CNRS/ INRIA, France 2 DCC – Federal University of Bahia, Brazil
OPODIS, Nice, Dec. 2013
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA
Roadmap
Context/Motivation
Model for dynamic network
Eventual election algorithm
Conclusion
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 2
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 3
Dynamic self-organized systems Wireless ad-hoc networks (WMN, WSN)
Key features Unknown membership Multi-hop networks Dynamic changing topology
churn, node mobility Transmission range communication
broadcast to neighborhood Asynchronous systems
no bound on processor speed and transmission delays
Context
Eventual election
The Ω failure detector satisfies (“eventual leader election”):
there is a time after which every correct process always trusts the same correct process
Ω is the weakest failure detector for solving Consensus in the crash failure model
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 4
Context
p1
p3
p4 p5
p2
Ω=p2
Ω=p2
Ω=p2
correct
crashed
Model for Failure Detection in Dynamic Networks
The membership is unknown A node is not aware about the set of nodes nor the number of them.
Finite arrival model The network is dynamic composed of infinite mobile nodes, but each run consist
of a finite set n nodes. The system is asynchronous
There are no assumptions on the relative speed of processes nor on message transfer delays.
Communication Channels are fair-lossy there is no message duplication, modification or creation
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 5
Note: node = process
Model (cont’d)
Communication graph is dynamic Dynamic topology represented by a time-varying graph
(TVG) [CFQS11] : TVG = <Nodes, Edges, Time, EdgePresence, Latency, NodePresence>
A node p can reach q if there exists a journey between p and q (i.e., a path over time between p and q)
p q
time 1 2 3
p r r s s qLatency 0.5 0.7 1
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 6
Processes status
Let knownq set denotes the partial knowledge of q.
2 sets of nodes :
STABLE (correct): nodes eventually and permanently correct
FAULTY: nodes which crash
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 7
Transmission range
mobile node
Stable node
Model
Network connectivity
Transmission TVG induced by the stable communicating nodes in the system A message sent at time t induces an edge at t in the
transmission TVG
Network recurrent connectivity TVG of class 5
There exists a journey between all stable nodes at any time Eventually, the transmission TVG is connected other time
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 8
Timer-based x Timer-free assumptions
The Ω failure can not be implemented in a pure asynchronous system Two approaches:
Timer-based Constraint to satisfy message transfer delays Channels are eventually timely
Message exchange pattern Constraint to satisfy a message delivery order Query-response mechanism
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 9
Properties to implement an
Stable Termination Property (SatP): Each QUERY must be received by at least one stable and
known node Necessary for the diffusion of the information
Stabilized Responsiveness Property (SRP): There exists a time t after which all nodes of pi 's
neighborhood receive, to every of their queries, a response from pi which is always among the first responses
SRP should be hold for at least one stable known node (the eventual leader)
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 10
An Eventual Leader Election Algorithm
Principle Election of a leader process based on punishment Periodic local query-response exchange
Wait for responses If pj, locally known by pi, does not respond to a query of p
among i, pj is punished by pi.
Exchange of information
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 11
r not punished
r punished
r not punished
p
qWaiting for q responses
…
r
Ne
igh
bo
rho
od
q
An Eventual Leader Election Algorithm
Notation (pi) midi: a counter to timestamp every query-response message;
local_knowni: the current knowledge of pi about its neighborhood set of tuples <midj,pj>; max known midj
global_knowni: the current knowledge of pi about the membership of the system set of tuples <midj,pj>; max known midj
punishi: punished processes by pi
set of tuples <ct,p>
Query and Response message from pi
<midi, punishi, global_knowni>
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 12
Leader Election: Sending of Query
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 13
punishment
* - pj is a neighbor of pi, - pj does not answer to pi, - pj is not suspected to have moved
*
Reception of Query and Response;Invocation of the Leader
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 14
*
*
*update of pi’s state about punishment, membership, and pi’s neighborhood with more recent information : keeps the tuples with the greatest counter.
*process with the smallest counter
*
15
Exemple: Mobility of nodes
1
2
3
<1,1>,<1,2>,<1,3>,<1,4>
<0,1>
<1,2>,<1,3>,<1,4>
4
local_known1
punished1
global_known1
<2,4>
<0,1>,<1,4> <0,1>,<2,4>
x:<x,4> in local_known1 < y:<y,4> in global_known1
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA
<0,1>,<3,4>
1 stops punishing 4
5
<1,2>,<1,3>,<2,4>
Conclusion
Model and and a timer-free algorithm to solve the eventual leader election in mobile dynamic systems
The algorithm implement the Ω class of failure detectors using: Query-response approach
Failure detection and exchange of information TGV framework
Dynamics of the Network
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 16
Thank you
OPODIS 2013: Paris 6 (CNRS/INRIA) and UFBA 17