Download presentation

Presentation is loading. Please wait.

Published byRuben Gritton Modified about 1 year ago

1
Stochastic optimization of service provision with selfish users C.F. Chiasserini, P. Giaccone, E.Leonardi Department of Electronics and Telecommunications F. Altarelli, A. Braunstein, L. Dall’Asta, R. Zecchina – Department of Applied Science and technology

2
Outline NETSTAT - Budapest Motivational scenario – WiFi green AP BP-based methodology Performance evaluation

3
Green AP NETSTAT - Budapest Scenario: – large WiFi network, with redundant coverage e.g., Politecnico campus network – protocol available to turn on/off APs e.g., Energy-wise protocol implemented in Cisco devices – large population of users, each with a given probability of being present Aim: – reduce power consumption by turning off some APs without affecting (with high probability) the minimum bandwidth of each users

4
Optimization problem NETSTAT - Budapest given for each user u – position (x u, y u,z u ) – probability of being present and active p u given the set of possible association rates from users to APs – r ua, between user u and its neighbouring AP a first criteria: maximize the number of APs to turn off subject to a minimum bandwidth guaranteed for each user second criteria: maximize the achievable bandwidth

5
Realistic scenario NETSTAT - Budapest Available data from the network administrators at Politecnico – full control of the WiFI network using Cisco proprietary solutions – position(x,y,floor) and connection log of each AP

6
Aps’ log NETSTAT - Budapest 20/06/2012, from 9:00 to 20:00 33 APs: for each AP, sampled every hour – AP MAC, number of associated clients, number of authenticated clients 1126 users: for each user – AP to which she is associated – association time interval – total data exchanged – average SNR/RSSI

7
User location and presence NETSTAT - Budapest Assumption: users are located at random around an AP Assumption: the presence probability p u for user u at time t is evaluated as: number of users connected at time t number of users connected in the whole day

8
Coverage graph NETSTAT - Budapest 20138

9
Rate model NETSTAT - Budapest given the distance between user u and AP a, we adopt an empirical multifloor propagation model validated in the literature for to evaluate the association rate of each user r ua the bandwidth among users is divided according to a standard model taking into account the different association rates and the protocol overheads

10
Methodology for the solver NETSTAT - Budapest use some classical iterative algorithm to turn OFF the APs – e.g. greedy decimation starting from all APs in ON state use belief propagation(BP) to evaluate efficiently the cost function

11
Problem definition NETSTAT - Budapest bipartite graph of users {u 1, …, u U } and APs {s 1, …, s S } – t u = 1 (present), 0 (absent), with probability p u – x s = 1 (AP on), 0 (AP off) – operational cost r s of AP s – w us = payoff of u selecting AP s – w su = load on AP s by user u – capacity c s = maximum load on AP s

12
Factor graph representation NETSTAT - Budapest Constraints: 1.User connect to at most one AP 2.Capacity constraints 3.Users maximize their payoff

13
Objective function NETSTAT - Budapest evaluation process of the cost function – fix t (user presence) selfish behavior of the users induces Nash Equilibrium Points (NEPs) average across all NEPs – average across all t novelty: use “mirror messages”

14
Validation NETSTAT - Budapest mirror approach vs. sampling of NEPs (4 AP, 12 users) S=number of istances of t (user presence)

15
Optimization result NETSTAT - Budapest results obtained by switching off the APs in Politecnico scenario

16
Conclusions NETSTAT - Budapest We propose an novel belief propagation approach to compute the costs of different service configurations – averaging across all the possible Nash Equilibrium Points – more efficient than Montecarlo approaches Useful for algorithm to solve stochastic allocation problems Proof of concept – green AP in a corporate WiFI network

Similar presentations

© 2016 SlidePlayer.com Inc.

All rights reserved.

Ads by Google