Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.
-
Upload
charles-wilkerson -
Category
Documents
-
view
216 -
download
0
Transcript of Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.
![Page 1: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/1.jpg)
Performance EvaluationPerformance Evaluation
Lecture 4: Epidemics
Giovanni Neglia
INRIA – EPI Maestro
12 January 2013
![Page 2: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/2.jpg)
Outline
Limit of Markovian modelsMean Field (or Fluid) models
• exact results• Extensions
- Epidemics on graphs- Reference: ch. 9 of Barrat,
Barthélemy, Vespignani “Dynamical Processes on Complex Networks”, Cambridge press
• Applications to networks
![Page 3: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/3.jpg)
SI on a graph
Susceptible
InfectedAt each time slot, each link outgoing from an infected node spreads the disease with probability pg
![Page 4: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/4.jpg)
Can we apply Mean Field theory?Formally not, because in a graph the
different nodes are not equivalent……but we are stubborn
![Page 5: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/5.jpg)
Derive a Mean Field modelConsider all the nodes equivalente.g. assume that at each slot the graph
changes, while keeping the average degree <d>• Starting from an empty network we add a link
with probability <d>/(N-1)
k=1
![Page 6: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/6.jpg)
Derive a Mean Field modelConsider all the nodes equivalente.g. assume that at each slot the graph
changes, while keeping the average degree <d>• Starting from an empty network we add a link
with probability <d>/(N-1)
k=2
![Page 7: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/7.jpg)
Derive a Mean Field modeli.e. at every slot we consider a sample of
an ER graph with N nodes and probability <d>/(N-1)• Starting from an empty network we add a link
with probability <d>/(N-1)
k=2
![Page 8: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/8.jpg)
Derive a Mean Field modelIf I(k)=I, the prob. that a given susceptible
node is infected is qI=1-(1-<d>/(N-1) pg)I
and (I(k+1)-I(k)|I(k)=I) =d Bin(N-I, qI)
k=2
![Page 9: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/9.jpg)
Derive a Mean Field modelIf I(k)=I, the prob. that a given susceptible
node is infected is qI=1-(1-<d>/(N-1) pg)I
and (I(k+1)-I(k)|I(k)=I) =d Bin(N-I, qI)• Equivalent to first SI model where p=<d>/(N-1)
pg
• We know that we need p(N)=p0/N2
i(N)(k) ≈ μ2 (kε(N))=1/((1/i0-1) exp(-k p0/N)+1)== 1/((1/i0-1) exp(-k <d> pg)+1)
• The percentage of infected nodes becomes significant after the outbreak time 1/(<d>pg)
How good is the approximation practically?• It depends on the graph!
![Page 10: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/10.jpg)
Let’s try on Erdös-Rényi graph Remark: in the calculations above we
had a different sample of an ER graph at each slot, in what follows we consider a single sample
![Page 11: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/11.jpg)
ER <d>=20, pg=0.1, 10 runs
N=20 N=200 N=2000
i(N)(k) ≈ 1/((1/i0-1) exp(-k <d> pg)+1)
![Page 12: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/12.jpg)
Lesson 1
System dynamics is more deterministic the larger the network is
For given <d> and pg, the MF solution shows the same relative error
![Page 13: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/13.jpg)
ER <d>=20, 10 runs
N=2000,pg=0.1
N=2000,pg=1/2000
![Page 14: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/14.jpg)
Lesson 2
For given <d>, the smaller the infection probability pg the better the MF approximation
– Why?
![Page 15: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/15.jpg)
Changing the degreeER N=1000, <d>pg=0.1, 10 runs
i(N)(k) ≈ 1/((1/i0-1) exp(-k <d> pg)+1)
Model<d>=10<d>=100
![Page 16: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/16.jpg)
Lesson 3
Given <d>pg, the more the graph is connected, the better the MF approximation Why?
![Page 17: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/17.jpg)
A different graph Ring(N,k)
![Page 18: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/18.jpg)
Ring vs ER, N=2000, <k>=10
ModelErdos-RenyiRing
![Page 19: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/19.jpg)
Lesson 4
The smaller the clustering coefficient, the better the MF approximation Why?
![Page 20: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/20.jpg)
Heterogeneous Networks
Denote P(d) the probability that a node has degree d
If the degree does not change much, we can replace d with <d>– what we have done for ER graphs (N,p)
• Binomial with parameters (N-1,p)
How should we proceed (more) correctly?– Split the nodes in degree classes– Write an equation for each class
Remark: following derivation will not be as rigorous as previous ones
![Page 21: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/21.jpg)
Heterogeneous Networks
Nd number of nodes with degree d (=N*P(d))
Id: number of infected nodes with degree d
Given node i with degree d and a link eij, what is the prob. that j has degree d’?– P(d’)? NO
and if degrees are uncorrelated? i.e. Prob(neighbour has degree d'|node has a degree d) independent from d,– P(d’)? NO– Is equal to d'/<d> P(d')
![Page 22: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/22.jpg)
Heterogeneous Networks
Given node i with degree d and a link eij Prob. that j has degree d’ is– d'/<d> P(d’)
Prob. that j has degree d’ and is infected
– d'/<d> P(d’) Id’/Nd’
– more correct (d’-1)/<d> P(d’) Id’/Nd’
Prob. that i is infected through link eij is
– p = pg Σd’ (d’-1)/<d> P(d’) Id’/Nd’
Prob. that i is infected through one link– 1-(1-p)d
![Page 23: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/23.jpg)
Heterogeneous Networks
E[(Id (k+1)-Id (k)|I (k)=I)] = (Nd-Id)(1-(1-p)d)
− p = pg Σd’ (d’-1)/<d> P(d’) Id’/Nd’
fd(N)(i)=(1-id)(1-(1-p)d)
− id = Id/Nd
− if we choose pg = pg0 /N
− fd(i)= pg0 (1-id) d Σd’(d’-1)/<d> P(d’) id’
did(t)/dt=fd(i(t))=pg0 (1-id(t)) d Θ(t)
Θ
![Page 24: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/24.jpg)
Heterogeneous Networks
did(t)/dt=fd(i(t))=pg0 (1-id(t)) d Θ(t),
− for d=1,2…
− Θ(t)=Σd’(d’-1)/<d> P(d’) id’(t)
− id(0)=id0, for d=1,2…
If id(0)<<1, for small t
− did(t)/dt ≈ pg0 d Θ(t)
− dΘ(t)/dt = Σd’(d’-1)/<d> P(d’) did’(t)/dt ≈ pg0 Σd’(d’-1)/<d> P(d’) d’ Θ(t) =
= pg0 (<d2> - <d>)/<d> Θ(t)
![Page 25: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/25.jpg)
Heterogeneous Networks
dΘ(t)/dt ≈ pg0(<d2>-<d>)/<d> Θ(t)
− Outbreak time: <d>/((<d2>-<d>) pg0)
• For ER <d2>=<d>(<d>+1), we find the previous result, 1/(<d>pg0)
• What about for Power-law graphs, P(d)~d-γ?
For the SIS model:
− dΘ(t)/d ≈ pg0(<d2>-<d>)/<d> Θ(t) – r0 Θ(t)
− Epidemic threshold: pg0 (<d2>-<d>)/(<d>r0)
![Page 26: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/26.jpg)
Outline
Limit of Markovian modelsMean Field (or Fluid) models
• exact results• extensions• Applications
- Bianchi’s model- Epidemic routing
![Page 27: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/27.jpg)
Decoupling assumption in Bianchi’s modelAssuming that retransmission processes
at different nodes are independent• Not true: if node i has a large backoff
window, it is likely that also other nodes have large backoff windows
We will provide hints about why it is possible to derive a Mean Field model…
then the decoupling assumption is guaranteed asymptotically
![Page 28: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/28.jpg)
References
Benaïm, Le Boudec, “A Class of Mean Field Interaction Models for Computer and Communication Systems”, LCA-Report-2008-010
Sharma, Ganesh, Key, “Performance Analysis of Contention Based Medium Access Control Protocols”, IEEE Trans. Info. Theory, 2009
Bordenave, McDonarl, Proutière, “Performance of random medium access control, an asymptotic approach”, Proc. ACM Sigmetrics 2008, 1-12, 2008
![Page 29: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/29.jpg)
Bianchi’s model
N nodes, K possible stages for each node, in
stage i (i=1,…V) the node transmit with probability q(N)
i (e.g. q(N)i =1/W(N)
i)If a node in stage i experiences a
collision, it moves to stage i+1 If a node transmits successfully, it
moves to stage 0
![Page 30: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/30.jpg)
Mean Field model
We need to scale the transmission probability: q(N)
i =qi/Nf(N)(m)=E[M(N)(k+1)-M(N)(k)|M(N)(k)=m]f1
(N) (m)=E[M1
(N)(k+1)-M1(N)(k)|M1
(N)(k)=m]
Pidle=Πi=1,…V(1-qi(N))m
iN
The number of nodes in stage 1 • increases by one if there is one successful
transmission by a node in stage i<>1• Decreases if a node in stage 1 experiences
a collision
![Page 31: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/31.jpg)
Mean field modelPidle=Πi=1,…V(1-qi
(N))miN -> exp(-Σiqi mi)
• Define τ(m)= Σiqi mi
The number of nodes in stage 1 • increases by one if there is one successful
transmission by a node in stage i<>1
- with prob. Σi>1 mi N qi(N) Pidle/(1-qi
(N))
• Decreases if a node in stage 1 experiences a collision
- with prob. m1 N q1(N) (1-Pidle/(1-q1
(N))
f1(N)
(m)=E[M1(N)(k+1)-M1
(N)(k)|M1(N)
(k)=m]== Σi>1miqi
(N)Pidle/(1-qi(N))
– m1q1(N)(1-Pidle/(1-q1
(N)))
![Page 32: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/32.jpg)
Mean field modelPidle=Πi=1,…V(1-qi
(N))miN -> exp(-Σiqi mi)
• Define τ(m)= Σiqi mi
f1(N)
(m)=Σi>1miqi(N)Pidle/(1-qi
(N))
– m1q1(N)(1-Pidle/(1-q1
(N)))
f1(N)
(m) ~ 1/N (Σi>1miqi e-τ(m)–m1q1(1-e-τ(m)))f1
(N) (m) vanishes and ε(N)=1/N,
continuously differentiable in m and in 1/NThis holds also for the other componentsNumber of transitions bounded=> We can apply the Theorem
![Page 33: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/33.jpg)
Outline
Limit of Markovian modelsMean Field (or Fluid) models
• exact results• extensions• Applications
- Bianchi’s model- Epidemic routing
![Page 34: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/34.jpg)
Mean fluid for Epidemic routing (and similar)1. Approximation: pairwise intermeeting
times modeled as independent exponential random variables
2. Markov models for epidemic routing3. Mean Fluid Models
![Page 35: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/35.jpg)
35
Inter-meeting times under random mobility (from Lucile Sassatelli’s course)
Inter-meeting times mobile/mobile have shown to follow an exponential distribution
[Groenevelt et al.: The message delay in mobile ad hoc networks. Performance Evalation, 2005]
Pr{ X = x } = μ exp(- μx)CDF: Pr{ X ≤ x } = 1 - exp(- μx), CCDF: Pr{ X > x } =
exp(- μx)
M. Ibrahim, Routing and Performance Evaluation of Disruption Tolerant Networks, PhD defense, UNS 2008
MA
ESTR
O
![Page 36: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/36.jpg)
Pairwise Inter-meeting time
ttTP
etTP t
)(log
)(
![Page 37: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/37.jpg)
Pairwise Inter-meeting time
ttTP
etTP t
)(log
)(
![Page 38: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/38.jpg)
Pairwise Inter-meeting time
ttTP
etTP t
)(log
)(
![Page 39: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/39.jpg)
2-hop routing
Model the number of occurrences of the message as an absorbing Continuous Time Markov Chain (C-MC):
• State i{1,…,N} represents the number of occurrences of the message in the network.
• State A represents the destination node receiving (a copy of) the message.
![Page 40: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/40.jpg)
Model the number of occurrences of the message as an absorbing C-MC:
• State i{1,…,N} represents the number of occurrences of the message in the network.
• State A represents the destination node receiving (a copy of) the message.
Epidemic routing
![Page 41: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/41.jpg)
41T. G. Kurtz, Solutions of Ordinary Differential Equations as Limits of Pure Jump Markov Processes, Journal of
Applied Probability, vol. 7, no. 1, pp. 49–58, 1970. M. Benaïm and J.-Y. Le Boudec, A class of mean field interaction models for computer and communication
systems, Performance Evaluation, vol. 65, no. 11-12, pp. 823–838, 2008.
(from Lucile Sassatelli’s course)
![Page 42: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/42.jpg)
42T. G. Kurtz, Solutions of Ordinary Differential Equations as Limits of Pure Jump Markov Processes, Journal of
Applied Probability, vol. 7, no. 1, pp. 49–58, 1970. M. Benaïm and J.-Y. Le Boudec, A class of mean field interaction models for computer and communication
systems, Performance Evaluation, vol. 65, no. 11-12, pp. 823–838, 2008.
(from Lucile Sassatelli’s course)
![Page 43: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/43.jpg)
43T. G. Kurtz, Solutions of Ordinary Differential Equations as Limits of Pure Jump Markov Processes, Journal of
Applied Probability, vol. 7, no. 1, pp. 49–58, 1970. M. Benaïm and J.-Y. Le Boudec, A class of mean field interaction models for computer and communication
systems, Performance Evaluation, vol. 65, no. 11-12, pp. 823–838, 2008.
(from Lucile Sassatelli’s course)
![Page 44: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/44.jpg)
44T. G. Kurtz, Solutions of Ordinary Differential Equations as Limits of Pure Jump Markov Processes, Journal of
Applied Probability, vol. 7, no. 1, pp. 49–58, 1970. M. Benaïm and J.-Y. Le Boudec, A class of mean field interaction models for computer and communication
systems, Performance Evaluation, vol. 65, no. 11-12, pp. 823–838, 2008.
(from Lucile Sassatelli’s course)
![Page 45: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/45.jpg)
45Zhang, X., Neglia, G., Kurose, J., Towsley, D.: Performance Modeling of Epidemic Routing. Computer Networks 51, 2867–2891 (2007)
Two-hopEpidemic
MAESTRO
(from Lucile Sassatelli’s course)
![Page 46: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/46.jpg)
A further issue
Model the number of occurrences of the message as an absorbing Continuous Time Markov Chain (C-MC):
• We need a different convergence result
[Kurtz70] Solution of ordinary differential equations as limits of pure jump markov processes, T. G. Kurtz, Journal of Applied Probabilities, pages 49-58, 1970
![Page 47: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/47.jpg)
[Kurtz1970]
{XN(t), N natural}a family of Markov process in Zm
with rates rN(k,k+h), k,h in Zm
It is called density dependent if it exists a continuous function f() in Rm such thatrN(k,k+h) = N f(1/N k, h), h<>0
Define F(x)=Σh h f(x,h)
Kurtz’s theorem determines when {XN(t)} are close to the solution of the differential equation:
![Page 48: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/48.jpg)
The formal result [Kurtz1970]Theorem. Suppose there is an open set E in Rm and a constant M such that |F(x)-F(y)|<M|x-y|, x,y in E supx in EΣh|h| f(x,h) <∞,
limd->∞supx in EΣ|h|>d|h| f(x,h) =0
Consider the set of processes in {XN(t)} such that limN->∞ 1/N XN(0) = x0 in Eand a solution of the differential equation
such that x(s) is in E for 0<=s<=t, then for each δ>0
)),(()(
sxFs
sx
0)()(1
supPrlim0
sXsXN N
tsN
![Page 49: Performance Evaluation Lecture 4: Epidemics Giovanni Neglia INRIA – EPI Maestro 12 January 2013.](https://reader036.fdocuments.us/reader036/viewer/2022062423/56649e6f5503460f94b6ca5f/html5/thumbnails/49.jpg)
Application to epidemic routingrN(nI)=λ nI (N-nI) = N (λN) (nI/N) (1-nI/N)
assuming β = λ N keeps constant (e.g. node density is constant)
f(x,h)=f(x)=x(1-x), F(x)=f(x)as N→∞, nI/N → i(t), s.t.
with initial condition
multiplying by N
))(1)(()(' tititi
Nni IN
/)0(lim)0(
))()(()(' tINtItI