Using a History Based Approach to Predict Topology Control - - PowerPoint PPT Presentation

Using a History‐Based Approach to Predict Topology Control Information in Mobile Ad Hoc Networks

7th Int. Conf. on Internet and Distributed Computing Systems, IDCS 2014 September 24th, 2014. Calabria, Italy Pere Millán1, C. Molina1, R. Meseguer2, S. F. Ochoa3, R. Santos4

1Universitat Rovira i Virgili, Tarragona, Spain 2Universitat Politècnica de Catalunya, Barcelona, Spain 3Universidad de Chile, Santiago, Chile 4Universidad Nacional del Sur, Bahia Blanca, Argentina

• Motivation
• Predicting Topology Control Information (TCI)
• Experimental Framework & Results
• Conclusions & Future Work

OLSR

Outline

2

Motivation

3

Motivation

Several social computing participation strategies use mobile ad hoc or opportunistic networks Several social computing participation strategies use mobile ad hoc or opportunistic networks

4

• Routing protocols in mobile collaboration scenarios

– Must be simple, efficient, reliable and quickly adapt to changes in the network topology – Should minimize delivery of topology control information (TCI) to avoid consuming too much devices’ energy

• Link‐state proactive‐routing protocols:

– Low latency (using an optimized and known data‐path ) – Cost: periodically flooding the network with TCI

Motivation

… and when the number of nodes is high … … and when the number of nodes is high …

5

… can overload the network!!! … can overload the network!!!

6

… can we address the problem of delivering much control information through the network? … can we address the problem of delivering much control information through the network?

7

– Extends the previous work on TCI prediction [6‐8]. – Uses a time window with historical node TCI info … … to predict next TCI. – Named “History‐Based Prediction” (HBP). – HBP Performance: determined by simulations in several mobile scenarios.

We present and evaluate a new strategy for predicting TCI in mobile ad hoc and opportunistic networks

8

Predicting TCI

9

• Idea:

– Use historical Topology Control Information (TCI) to make predictions of the next control packets (CP).

• Questions to answer:

– What performance and limits has this approach? – In which mobile computing scenarios this proposal can provide a real benefit?

Predicting TCI using Past Information

10

• Each node keeps updated locally (in a table)

the recent TCI history received from its neighbors.

• Prediction at each node:

– Input: recent TCI history. – Output: a prediction of TCI for each neighbor (guess network topology without delivering control info).

• Prediction can be done when previous TCI received

matches TCI previously stored.

• HBP predicts a state already appeared in the past.

HBP Assumptions

11

• Control Packets sequence:

AAAABABAACBAABBAB

• Table contents

(patterns with 2 control packets):

HBP table example

Pattern Next Count Last AA A B C 2 2 1 # AB A B 2 1 # AC B 1 # BA A B 2 2 # BB A 1 # CB A 1 #

12

• Unbounded

– More flexibility to identify movement patterns.

• One table per node.
• Movement pattern (stored in the table):

– Sequence of 1+ TCI packets seen in the past.

• Attached to every pattern stored in the table:

– A list of all packets appeared after each pattern. – Statistical information: last packet, most frequent.

HBP tables with historic information

13

• NS‐3 (4 hours) + BonnMotion.
• Mobility: Random Walk, Nomadic, SLAW.
• OLSR protocol (HELLO: 2 s / TC: 3 s).
• 300x300 m open area (beach, park). Free to move/interact.
• Node devices:
• All similar (capabilities ≈ iPhone 4).
• Wi‐Fi (detect others & Exchange CI). Range: 80 m / BW ≥ 50 kbps.
• 10, 20, 30, 40 nodes, randomly deployed.
• 1 m/s (walking), 2 m/s (trotting), 4 m/s (running), and 6 m/s (bicycling).

Experimental Framework

14

Results

15

• To help us understand predictability

and prediction opportunity limits of our proposal.

• Maximum reachable prediction accuracy:

– Count if a certain TCI packet has ever appeared in the past. – If it appeared once, we assume it could be predicted.

We quantify TCI repetition over time

16

Results: Predictability Limits

%TCI packets appeared in the past

• 3 mobility models (1 m/s)
• 10‐40 nodes density

%TCI packets appeared in the past

• 3 mobility models (1 m/s)
• 10‐40 nodes density

About 80% for 10 nodes High prediction potential About 80% for 10 nodes High prediction potential Prediction capability does not depend on mobility model Prediction capability does not depend on mobility model Prediction limits decrease when node density increases. Prediction limits decrease when node density increases.

17

• We also analyze

the representativeness of the most‐frequent packets, with respect to the whole set

• f packets received by a node over time.
• This will give us:

– A first understanding about how difficult is to make right predictions. – And which is the amount of historical data that must be tracked to make these predictions.

Packet representativeness

18

Frequency of Observed Control Packets

What control packets appear most frequently? What control packets appear most frequently?

30% of control packets represent 70% total observed

A small subset of packets represent the most delivered A small subset of packets represent the most delivered

Does not depend

• n node density

nor mobility models Does not depend

• n node density

nor mobility models

Many opportunities to predict with a small subset of packets. Many opportunities to predict with a small subset of packets.

19

• In case of wrong predictions (miss), classification:

– It could be correctly predicted, if the right control packet was in the list of this pattern (missPred). – If not, it could not be predicted (missNoPred).

• This identifies the limits of HBP

and how far an approach is from the best.

Types of wrong predictions

20

• History Depth (HD) metric:

– Number of TCI packets in the movement patterns.

• HD range considered:

– 0 to 5 TCI packets.

• High HD values (long sequences):

– More accurate predictions, few opportunities to predict.

• Low HD values (short sequences):

– Less accurate predictions, more opportunities to predict.

History Depth

21

History‐based prediction (varying HD)

22

HD

HD=0: largest %hits HD=0: largest %hits But important %misses too But important %misses too

Large HD: ‐ Predictions + Accurate. Large HD: ‐ Predictions + Accurate. noPred increases with number of nodes (predictability limits) noPred increases with number of nodes (predictability limits)

• Last value  last packet seen after this pattern.
• Most‐frequent  highest count packet.
• History‐based Random  any past packet

seen after this pattern.

• Previous example:

AAAABABAACBAABBAB

HBP flavors

Pattern Next Count Last AA A B C 2 2 1 # AB A B 2 1 # AC B 1 # BA A B 2 2 # BB A 1 # CB A 1 #

Last value: B Last value: B Most frequent: A Most frequent: A History‐based random: any of A/B History‐based random: any of A/B Pure random: any of A/B/C Pure random: any of A/B/C

23

?

History‐based prediction (different policies)

24

Always predicts (wrongly) Always predicts (wrongly) Much better results when using history (even random) Much better results when using history (even random)

History information provides more accurate predictions. History information provides more accurate predictions.

History‐based prediction (different mobility)

25

Similar behavior (difference <10%) Similar behavior (difference <10%) Mobility models do not present significant differences in prediction capability. Mobility models do not present significant differences in prediction capability.

• HBP flavors must succeed in the predictions

but also not predict when success is not guaranteed.

– Success reduces network traffic and saves energy. – Wrong predictions can skew the network topology map and decrease the reliability of the process.

• We include a confidence mechanism to determine

the likelihood that a prediction is correct.

– Aim: maximize right and minimize wrong predictions.

Prediction confidence

26

• Simple confidence mechanism for HBP:

– Saturated counter for each pattern in history table. – 2‐bit counter (values range: 0 to 3). – Counter incremented when the prediction is right. – Counter decremented when the prediction is wrong. – Prediction is confident when counter ≥ 2. – Counter initialized as 1 (no confidence).

Implementing confidence

27

History‐based prediction (2‐bit confidence)

28

Our goal: Maximize noConf/miss Minimize noConf/hit Our goal: Maximize noConf/miss Minimize noConf/hit Using confidence: Less predictions (mainly hits, few misses) Using confidence: Less predictions (mainly hits, few misses)

Using a confidence mechanism we can minimize prediction errors. Using a confidence mechanism we can minimize prediction errors.

• Previous HBP flavors use fixed History Depth (HD).
• We analyze an additional HBP flavor

where History Depth is dynamic (prediction tree):

– Start the prediction with the largest HD. – If prediction is not possible (not confident or missing movement pattern), decrease HD value (shorter pattern), and repeat. – Repeat until prediction is possible or HD reaches 0.

Dynamic history depth (tree)

29

Fixed History‐Depth vs. Dynamic (tree)

30

Tree minimizes noPred Tree minimizes noPred But increases significantly total hits But increases significantly total hits … decreases hits+misses … decreases hits+misses

Confidence mechanism + tree = Better results: total hits maximized, few misses. Confidence mechanism + tree = Better results: total hits maximized, few misses.

31

Conclusions Future Work

Reduces network traffic and saves energy 50%‐80% of packets appeared in the past, HBP upper limits are high for many scenarios Few packets contribute to total packets (high opportunity to predict TCI) At least 30% of correct predictions in a worst‐case scenario (many nodes)

OLSR

History‐based Prediction: Conclusions

32

• OLSR with prediction [6]
• Assume the last TCI send by a node

will probably be repeated during the next round

• f information delivery (Last value, HD=0)

– Less amount of control packets transmitted – Saves computational processing and energy – Independent of the OLSR configuration – Self‐adapts to network changes – Medium density scenario: 57% control packets reduction, 60% energy reduction, with same OLSR performance

OLSRp

33

OLSR

Future Work

1) Analyze in detail all combinations of work scenarios ‐ Considering node density, speed, and mobility patterns 2) Develop more complex confidence mechanisms ‐ and combine prediction approaches ‐ Their benefits can be accumulated? 3) Analyze prediction performance in opp networks involving heterogeneous environments ‐ To address IoT‐based solutions.

34

Questions?

Thanks for Your Attention

7th Int. Conf. on Internet and Distributed Computing Systems, IDCS 2014 September 24th, 2014. Calabria, Italy

Questions?

7th Int. Conf. on Internet and Distributed Computing Systems, IDCS 2014 September 24th, 2014. Calabria, Italy