DYNAMIC CONTROL OF NETWORK PROTOCOLS

A New Vision for Future Self-organising Networks

Sven Tomforde, Emre Cakar and J

org H

ahner

Institute of Systems Engineering, Leibniz Universit

at Hannover, Appelstr. 4, 30167 Hannover, Germany

Keywords:

Organic Computing, Data communication network, Protocol, Parameter, Optimisation.

Abstract:

In recent years communication protocols have shown an increasing complexity, in particular in terms of the

number of variable parameters. Data communication networks like, e. g. , the Internet reach the limits of

their extensibility which leads to initiatives coping with the future of the Internet and data communication in

general. A ﬁrst step towards creating a sustainable solution without exchanging the whole system is to make

the static character of network protocols more ﬂexible. An adaptive behaviour of nodes within a network and

an autonomous, self-organising concept for their control strategies leads to a possible increase in performance

accompanied by an increase of extensibility. This paper presents a new vision of how to establish these

new control strategies mostly independent of the particular protocol by using the concepts of Organic and

Autonomic Computing. We introduce an adaptive and automated network control system for the dynamic and

self-organised control of protocol parameters. This system consists of two sub-systems: an on-line adaptation

mechanism and an off-line learning component. The current status is introduced in combination with the

deﬁnition of further challenges and ﬁelds of research.

1 INTRODUCTION

Recent years were characterised by a dramatical

growth of communication need and increase of traf-

ﬁc over data communication networks. In combi-

nation with the ascending number of protocols and

their varying conﬁguration possibilities (conﬁgura-

tion space) the complexity of the control task at each

node in the network is growing. Based on this ob-

servation, more and more often the question arises

whether the current structure of the network (in partic-

ular the Internet) will be able to cope with the increas-

ing demand (cf. e. g. (Handley, 2006)). This leads to

a new vision for the future of the Internet ((Siekkinen

et al., 2007)).

The number of researchers formulating the need to

exchange the complete set of techniques (e. g. proto-

cols, structure, etc.) is increasing steadily. One major

problem focused here is that the existing protocols are

designed as static solutions. Although the situation at

particular nodes within the communication networks

(in terms of e. g. resource usage, available bandwidth,

currently known neighbours, etc.) changes over time,

the conﬁguration is typically not adapted to the cur-

rent requirements. A possibility to solve this prob-

lem by keeping downward compatibility (this means

cooperation of static and dynamic solutions) is pre-

sented within this paper.

Based on the approaches of Organic Computing

(OC - cf. (Schmeck, 2005)) and Autonomic Comput-

ing (AC - cf. (Kephart and Chess, 2003)) an adaptive

network control system is introduced which aims at

coping with the large conﬁguration space. The sys-

tem is locally organised, adaptive, and has learning

abilities guaranteeing the best possible performance

for each node.

This paper presents an adaptive and automated

system for the dynamic and self-organised control of

network protocol parameters (e. g. values for time-

outs, maximum number of re-transmissions, number

of open connections, etc.). Section 2 focuses on

a short overview of already introduced approaches

to optimise network settings and concludes with the

statement that no system with the required properties

exists yet. In Section 3 the architecture and general

approach for the proposed system are described com-

bined with details on the technical realisation. After-

wards, Section 4 deﬁnes the challenges for the fur-

ther development and the research to be done until a

real-time operation can be applied. Finally, Section 5

concludes the vision for dynamic network control.

285

Tomforde S., Cakar E. and Hähner J. (2009).

DYNAMIC CONTROL OF NETWORK PROTOCOLS - A New Vision for Future Self-organising Networks.

In Proceedings of the 6th International Conference on Informatics in Control, Automation and Robotics - Intelligent Control Systems and Optimization,

pages 285-290

DOI: 10.5220/0002243302850290

 SciTePress

2 STATE OF THE ART

The dynamic selection of network protocol parame-

ter settings depends on the situation-based generation

of these settings. Therefore, the task can be divided

into two different subtasks: the off-line optimisation

of parameter settings for a given (observed) situation

and the on-line adaptation of the network controller

settings with a suitable parameter set.

The optimisation of parameter settings deals with

the problem to determine a set of parameters for a

given protocol that is as close to the optimum as possi-

ble. The task is characterised by the required amount

of time and the quality of the solution to be found.

In this context off-line means evaluating new possible

settings using simulation and thus without interfering

with the live-system.

There are several examples where authors opti-

mised the settings of their particular protocols, but

their intention has been to optimise a speciﬁc proto-

col and not to create a generic system. Considering

the techniques used in our system, the approach of

Montana and Redi is connected as they also use an

Evolutionary Algorithm (EA) to optimise a full cus-

tom communication protocol for military MANETs

(Montana and Redi, 2005). A similar optimisation of

a protocol (for underwater communications) using an

EA is described by S

ozer et al. (S

ozer et al., 2000).

Turgut et al. discuss the usage of a Genetic Algo-

rithm to optimise their MANET-based clustering pro-

tocol in (Turgut et al., 2002). They all compare their

achieved results to a manual optimisation. In contrast

to the network control system presented in this paper

the approaches are speciﬁc to the particular protocols,

but do not aim at providing a generic system which is

adaptable to different protocol types.

Due to the time-intensive process of generat-

ing optimal parameter sets an on-line usage in live-

systems is not applicable. Hence, such a solution

has to somehow combine the strengths of optimisa-

tion techniques with approaches to immediately react

on an observed stimulus. Although research commu-

nities are aware of the demand and it already has been

part of the vision of initiatives (Kephart and Chess,

2003) a solution has not been presented yet.

One approach towards a possible solution has

been described by Ye and Kalyanaraman (Ye and

Kalyanaraman, 2001). They introduced an adaptive

random search algorithm, which tries to combine the

stochastic advantages of pure random search algo-

rithms with threshold-based knowledge about extend-

ing the search. Their approach is based on the initial

system as presented in (Ye et al., 2001). In contrast to

our approach, Ye et al. propose a centralised system

that tackles the optimisation task for each node. To

allow for such a division of work between a central

server and the particular network nodes they have to

deal with problems like e. g. bandwidth usage, single

point of failure, or local knowledge accessible from

server-side.

3 SYSTEM

The motivation to develop a dynamically adapting

system has been formulated before (cf. AC (Kephart

and Chess, 2003) or Autonomic Networking (Jen-

nings et al., 2007)), a proof of concept is still miss-

ing. The system presented in this paper is a ﬁrst step

towards a possible realisation. Based upon our archi-

tecture as pictured in Fig. 1 and initially presented

in (Tomforde et al., 2009) the responsibilities of pa-

rameter set generation and on-line adaptation of the

control system are assigned to different layers. One

component (Layer 2) evolves new parameter sets not

being restricted by real-time requirements for time

and computation power. The other part reacts on

changing stimuli (observed situation). This division

of functionalities leads to the possibility that for an

observed situation no matching optimised parameter

set is available. In this case a covering mechanism

has to cope with the situation which chooses the best

possible control and adaptation strategy.

Technically, the architecture is realised by two

connected techniques. As described in e. g. (Schmid

et al., 2006), a combination of evolution and learn-

ing seems to be a promising solution to realise dy-

namic system-adaptations. Based on this assumption,

for the off-line part an Evolutionary Algorithm (EA)

is used in combination with a standard network simu-

lation tool. The on-line learning mechanism is based

on a modiﬁed version of Wilson’s Learning Classiﬁer

System XCS (Wilson, 1995).

Within this Section the goals of the system are de-

ﬁned, followed by a short introduction of the basic

concepts. The main part is dealing with the architec-

ture and the current status.

3.1 Goal Deﬁnition

The network control system as presented in this pa-

per aims at increasing the performance of data com-

munication. It allows for the dynamic adaptation of

network protocols to a continuously changing envi-

ronment. Based on the initially introduced concept

(Tomforde et al., 2009), the system requires organic

(in terms of OC) characteristics – a decentral, self-

organised approach leads to a stable, reliable con-

ICINCO 2009 - 6th International Conference on Informatics in Control, Automation and Robotics

286

trol, the system is able to learn and optimise its be-

haviour autonomously. The network control system

is generic, which means the controlled network proto-

col client can be exchanged. It supports a large set of

different protocol types (Peer-to-Peer, mobile ad-hoc,

wire-based, sensor, etc.) and protocols (e. g. BitTor-

rent (Cohen, 2003), Hyper-Gossiping (Khelil et al.,

2007), etc.). If possible, the autonomous network

control systems can collaborate and fullﬁll system-

wide goals by using local interactions. The goal deﬁ-

nition has some similarities with the manifest of Au-

tonomic Computing (Kephart and Chess, 2003) and

the Autonomic Management of Networks approach

as presented in (Jennings et al., 2007). In contrast to

the network-wide approach (which is unfeasible for

large networks like the Internet), we assume that a lo-

cally organised solution based on local rules and local

interactions will converge to a system-wide optimisa-

tion in most of the cases and therefore does not need

global knowledge and global control.

3.2 Basic Concepts

The architecture is based on two already known ap-

proaches: the Generic Observer/Controller Architec-

ture (Richter et al., 2006) and the 2-layered Archi-

tecture of the Organic Trafﬁc Control (OTC) Sys-

tem (Prothmann et al., 2008). The generic archi-

tecture describes an approach, where a decentralised

system (System under Observation/Control (SuOC))

is wrapped with an additional surveillance and feed-

back mechanism. Sensors and actuators are used to

monitor and control the SuOC by establishing a con-

trol loop. This control loop observes the behaviour of

the SuOC through sensors, compares the results with

expected behaviour and the current goals of the sys-

tem, decides what action is necessary and controls the

SuOC with the best known action through actuators.

Additionally, a memory function keeps track of his-

torical situations and control actions to be able to op-

timise the behaviour from existing knowledge.

The architecture of the OTC system is based on

this approach. The realisation of a real-world sce-

nario (control of trafﬁc lights at urban intersections)

leads to some restrictions – the main aspect is that the

system has to use only parameter sets with guaran-

teed performance. Therefore, the situation-dependent

creation of new parameter sets has been assigned to a

new layer within the architecture where a simulation-

based approach is performed off-line. This concept

can be also found in the network control system, but

the different domain leads to some modiﬁcations and

changes, which will be explained in the remainder of

this section.

3.3 Architecture

Similar to the 2-layered Architecture of the Organic

Trafﬁc Control (OTC) System (Prothmann et al.,

2008) our architecture consists of three parts: the

SuOC at Layer 0, an on-line adaptation mechanism

at Layer 1, and an off-line learning component at

Layer 2 (see Fig. 1). All three layers will be presented

in the following.

Figure 1: System architecture.

Layer 0: System under Observation and Control

The SuOC is a parametrisable Network Controller.

Due to the generic concept of the proposed system

it is not restricted to a particular set of protocols –

the only restriction applied is that it has to provide a

set of variable parameters and a local quality criterion

(e. g. duration of a download in Peer-to-Peer systems

or a weighted trade-off between energy consumption

and broadcast-covering for MANETs) for the perfor-

mance measurement. This means, protocols on all

layers (media access to application) can be controlled

in the same way as e. g. wire-based protocols or mo-

bile ad hoc networks. A good setup of the variable

parameters that match the current condition at the net-

work node has an important inﬂuence on the resulting

performance for these systems. In the architecture,

the parameter setup is optimised on-line by the O/C

component in Layer 1.

Layer 1: On-line Adaptation

The Layer 1 component can be divided into two dif-

ferent parts: an Observer and a Controller. The Ob-

DYNAMIC CONTROL OF NETWORK PROTOCOLS - A New Vision for Future Self-organising Networks

287

server is responsible for monitoring the situation at a

particular node. It measures those attributes having

inﬂuence on the selection of appropriate parameters

for the control strategy. This selection depends on

the speciﬁc controlled protocol and typically contains

attributes like buffer sizes, delay times, etc. Addi-

tionally, protocol-speciﬁc parameters like e. g. num-

ber of nodes in sending distance for MANET proto-

cols or available system resources like CPU, upload-

bandwidth, download-bandwidth, etc. for P2P pro-

tocols can be taken into account. Afterwards, these

values are aggregated to an abstract situation descrip-

tion realised as an n-dimensional vector with n equal

to the number of observed values.

The main part of the Controller is a Learning Clas-

siﬁer System, which is based on Wilson’s XCS as in-

troduced in (Wilson, 1995). The LCS maps the aggre-

gated input-information from the Observer to a rule

base of possible actions, the process is realised in ac-

cordance with Wilson’s approach. The basic change

in concept is that our LCS version is not able to cre-

ate new rules as this process can lead to unwanted

behaviour (random rule generation). This leads to

the problem that the system might not have a match-

ing rule for the currently observed situation, although

the system detects the demand to adapt the network

client. Therefore, a covering mechanism is needed,

which chooses the best possible action.

This covering is realised based on the assumption

that a classiﬁer whose condition part is located close

to the current situation description – although it does

not match it – is better than any other one existing

within the rule set. Due to this assumption a covering

process is executed which selects the ”nearest” classi-

ﬁer in terms of the Euclidian Distance calculated for

the n dimensional vectors (equal to the situation de-

scription) and using the centroids of the intervals used

for each interval predicate. This classiﬁer is copied,

its condition part is adapted to the current situation

description (using a standard interval size around the

given situation), and it is added to the rule set. Based

on this simple process we ensure to only use tested ac-

tions and we also ensure that at least one rule is con-

tained in the match set. Further details on the process

can be found in (Tomforde et al., 2009).

Layer 2: Off-line Learning

The existing set of classiﬁers and consequently the set

of existing parameter sets has to be extended for sit-

uations where no classiﬁer matches. This means, the

system has to be able to autonomously learn param-

eter sets for unforeseen situations. Within our archi-

tecture, the Layer 2-component is responsible for this

task. This component consists of three parts: an Ob-

server, an EA and a simulator. The Observer is re-

sponsible for capturing the current situation descrip-

tion provided by the Observer on Layer 1. As the cur-

rent usage of system components (in terms of CPU,

RAM, etc.) has inﬂuence on the selection process of

the LCS, the Observer is responsible for the schedul-

ing of optimisation tasks.

The EA is responsible for evolving new classiﬁers.

The algorithm is implemented as a standard Genetic

Algorithm (cf. e. g. (B

ack and Schwefel, 1996) for

details). This algorithm needs a possibility to analyse

the performance of the current parameter set, which

is done by using the standard network simulation tool

NS/2 (Web, 2009). The simulator needs a scenario

and an implementation of the current protocol. The

implementation is mandatory, but the conﬁguration

of the simulator depends on the observed situation as

measured by the Observer on Layer 1. Therefore, a

scenario is computed taking into account all observed

attributes (e. g. for BitTorrent: number of peers, seeds,

download and upload speeds, etc.)

4 RESEARCH ROADMAP

Within the previous Section the architecture of the

system has been described. The system based on this

has been realised and applied to a ﬁrst protocol (Bit-

Torrent - cf. (Cohen, 2003)). We demonstrated the

potential of our system and validated the feasibility

of our approach for a BitTorrent-based test scenario,

leading to an increase in terms of the objective func-

tion (amount of downloaded data or download-time)

of up to 20% (Tomforde et al., 2009). Further eval-

uation of this protocol is in progress. Additionally,

we are working on demonstrating the applicability of

our approach to other systems by replacing the SuOC

(adapt and optimise mobile ad-hoc network protocols

instead of a BitTorrent Client).

To completely achieve the goal as deﬁned in Sec-

tion 3.1 essential parts are still not investigated. The

following part of this Section will emphasise the main

focus of the future research based on this system.

Therefore, we introduce the main research topics in

accordance to the particular layers of our architecture.

4.1 Layer 0

The system aims at being generic in terms of control-

ling different protocols and protocol types. These pro-

tocols are situated at Layer 0 of our architecture as de-

picted in Fig. 1. To demonstrate the generic character

of our approach we are going to apply the system to

exemplary representatives of different protocol types.

ICINCO 2009 - 6th International Conference on Informatics in Control, Automation and Robotics

288

Figure 2: Collaborating network control systems.

Starting with BitTorrent as representative for Peer-

to-Peer systems and the current application to mo-

bile ad-hoc networks we will investigate the control

of other protocol domains. Therefore, protocols for

sensor networks are from interest as well as classical

Internet protocols (like TCP/IP) and very specialised

approaches like e. g. protocols for the communication

in smart camera networks (Hoffmann et al., 2008).

In addition to the application of different proto-

col types, we aim at extending the control scope to

cross-layer optimisation (Wang et al., 2005). This

means e. g. for TCP/IP that the conﬁguration of IP

is selected depending on the current situation and of

TCP depending on the conﬁguration of IP.

4.2 Layer 1

The performance of the on-line adaptation mecha-

nism at Layer 1 depends primarily on the applied

learning technique. Due to this dependency we aim

at validating the usage of our LCS by comparing it to

other learning techniques. Therefore, existing tech-

niques will be analysed based on the usage within our

architecture and implemented if promising.

In addition to the increase of performance by

analysing the learning component, the overall perfor-

mance can be increased by allowing for collaboration.

Neighboured entities should get the ability to collab-

orate with each other (see Fig. 2) in order to schedule

Layer 2 tasks, exchange knowledge, and avoid redun-

dant simulation-based learning.

4.3 Layer 2

As the off-line generation of new parameter sets is

resource- and time-consuming, an improvement is

necessary. The urgent target for this component is to

speed up the rule creation process. The approach to

solve this problem consists of two different strategies:

a speed-up at start time and an approximation at run-

time. At start time the system does not have any other

rules than the standard parameter set of the protocol,

which leads to the need of a fast mechanism to learn

rules for a set of exemplary rules within the conﬁgu-

ration space. These rules might have a lower quality

than an optimised one, but they will be replaced with

an optimised version during runtime.

The other aspect of the speed-up process at

Layer 2 is, that nodes might not have sufﬁcient re-

sources for the optimisation task (e. g. sensor net-

works). Hence, research here will focus on approx-

imating a reasonable parameter set (taking those pa-

rameter sets into account which are situated close to

the situation description). This means, an intelligent

inter- and extrapolation mechanism is needed in com-

bination with a half-centralised solution (e. g. periodic

updates from a service running on other nodes).

Another aspect of the Layer 2 optimisation is to

use stand-by time (no active optimisation task) to op-

timise the coverage of the conﬁguration space. This

means some kind of active learning may be used

to pro-actively generate parameter sets for situations

where currently no adequate parameter set is known.

Additionally, a collaboration mechanism will be help-

ful to schedule these active learning tasks for a set of

neighboured entities.

DYNAMIC CONTROL OF NETWORK PROTOCOLS - A New Vision for Future Self-organising Networks

289

5 CONCLUSIONS

This paper presented a system for the dynamic adap-

tation of network protocol parameters. The system

monitors the situation at particular nodes and reacts

on changes by adapting the communication protocol

client. It is able to learn new control strategies and

works on a self-organised basis. We explained our

position that the presented system will be able to in-

crease the performance of future communication net-

works without changing the whole technical back-

ground. Finally, we named the main research ﬁelds

for our approach based on the introduced goal. Our

system can also serve as a good testbed for the inves-

tigation of innate aspects of OC systems like trustwor-

thiness or collaboration patterns.

REFERENCES

ack, T. and Schwefel, H.-P. (1996). Evolutionary comput-

ing: An overview. In Proceedings of IEEE Conference

of Evolutionary Computing.

Cohen, B. (2003). Incentives Build Robustness in BitTor-

rent. In Proceedings of the 1st Workshop on Eco-

nomics of Peer-to-Peer Systems, Berkeley.

Handley, M. (2006). Why the internet only just works. BT

Technology Journal, 24(3):119–129.

Hoffmann, M., Wittke, M., Bernard, Y., Soleymani, R.,

and Hahner, J. (2008). Dmctrac: Distributed multi

camera tracking. Distributed Smart Cameras, 2008.

ICDSC 2008. Second ACM/IEEE International Con-

ference on, pages 1–10.

Jennings, B., van der Meer, S., Balasubramaniam, S.,

Botvich, D., Foghlu, M. O., Donnelly, W., and Strass-

ner, J. (2007). Towards autonomic management of

communications networks. Communications Maga-

zine, IEEE, 45(10):112–121.

Kephart, J. O. and Chess, D. M. (2003). The Vision of

Autonomic Computing. IEEE Computer, 36(1):41–

50.

Khelil, A., Marron, P. J., Becker, C., and Rothermel, K.

(2007). Hypergossiping: A generalized broadcast

strategy for mobile ad hoc networks. Ad Hoc Net-

works, 5:531–546.

Montana, D. and Redi, J. (2005). Optimizing parameters

of a mobile ad hoc network protocol with a genetic

algorithm. In GECCO ’05: Proc. of the 2005 confer-

ence on Genetic and evolutionary computation, pages

1993–1998, New York, NY, USA. ACM.

Prothmann, H., Rochner, F., Tomforde, S., Branke, J.,

uller-Schloer, C., and Schmeck, H. (2008). Organic

control of trafﬁc lights. In Proc. of the 5th Intern.

Conference on Autonomic and Trusted Computing.

Richter, U., Mnif, M., Branke, J., M

uller-Schloer, C.,

and Schmeck, H. (2006). Towards a generic ob-

server/controller architecture for Organic Computing.

In Hochberger, C. and Liskowsky, R., editors, INFOR-

MATIK 2006 – Informatik f

ur Menschen!, volume P-

93 of GI-Edition – Lecture Notes in Informatics (LNI),

pages 112–119. K

ollen Verlag.

Schmeck, H. (2005). Organic Computing – A new vi-

sion for distributed embedded systems. In Pro-

ceedings of the 8th IEEE International Symposium

on Object-Oriented Real-Time Distributed Computing

(ISORC’05), pages 201–203.

Schmid, S., Sifalakis, M., and Hutchison, D. (2006). To-

wards autonomic networks. In 3rd Intern. Annual

Conference on Autonomic Networking, Autonomic

Communication Workshop (IFIP), Lecture Notes in

Computer Science. Springer Verlag, Heidelberg.

Siekkinen, M., Goebel, V., Plagemann, T., Skevik, K.-A.,

Banﬁeld, M., and Brusic, I. (2007). Beyond the fu-

ture internet–requirements of autonomic networking

architectures to address long term future networking

challenges. Future Trends of Distributed Computing

Systems, IEEE International Workshop, 0:89–98.

ozer, E. M., Stojanovic, M., and Proakis, J. G. (2000). Ini-

tialization and routing optimization for ad-hoc under-

water acoustic networks. In Proc. of Opnetwork’00.

Tomforde, S., Steffen, M., H

ahner, J., and M

uller-Schloer,

C. (2009). Towards an organic network control sys-

tem. submitted for publication.

Turgut, D., Daz, S., Elmasri, R., and Turgut, B. (2002).

Optimizing clustering algorithm in mobile ad hoc net-

works using genetic algorithmic approach. In Proc.

of the IEEE Global Telecommunications Conference

(GLOBECOM ’02), pages 62 – 66.

Wang, J., Li, L., Low, S. H., and Doyle, J. C. (2005). Cross-

layer optimization in TCP/IP networks. IEEE/ACM

Trans. Netw., 13(3):582–595.

Web (2009). The Network Simulator - NS/2.

http://www.isi.edu/nsnam/ns/.

Wilson, S. W. (1995). Classiﬁer ﬁtness based on accuracy.

Evolutionary Computation, 3(2):149–175.

Ye, T., Harrison, D., Mo, B., Sikdar, B., Kaur, H. T., Kalya-

naraman, S., Szymanski, B., and Vastola, K. (2001).

Network Management and Control Using Collabora-

tive On-line Simulation. In Proceedings of IEEE ICC,

Helsinki, Finland. IEEE.

Ye, T. and Kalyanaraman, S. (2001). An adaptive random

search algorithm for optimizing network protocol pa-

rameters. Technical report, Rensselaer Polytechnic

Inst.

ICINCO 2009 - 6th International Conference on Informatics in Control, Automation and Robotics

290