USING ROUTING AGENTS FOR IMPROVING THE QUALITY

OF SERVICE IN GENERAL PURPOSE NETWORKS

Mohammad Mottaghi

and Masoud Daneshtalab

Dept. of Electrical and Computer Engineering, University of Tehran, Tehran, Iran

Dept. of Information Technology, Turku University, Turku, Finland

Keywords: Agent, Context-aware, Collaboration, Emergence, Routing, Network capacity, Utilization, QoS.

Abstract: In this paper, we have proposed a new routing method for general purpose networks. Each router in this

model is a context-aware agent and therefore the whole network of routers forms a community (society) of

context-aware agents which are ever learning and adapting the routing network. During routing each agent

keeps learning and specializing in routing some specific traffic type (e.g. intermittent audio packets) and

makes its neighboring routes optimum for this type of traffic whereas other agents are experts in other

traffic types. All agents are aware from their colleagues' expertise and when a new traffic type enters the

network all agents try to collaboratively recognize (or detect) its type and route it according to their past

experience (which is already learned). If the traffic type is unknown to all agents, one of them tries to learn

how to route it such that its QoS constraints are met better. The idea behind the proposed model is to

temporarily modify some routes whenever QoS constraints cannot be met in the network. Simulation results

show that the proposed model can improve QoS of the network by 12%.

1 INTRODUCTION

With the advent of the Internet many applications

emerged which wanted to communicate through the

Internet. Few years after the Internet became popular

various traffic types were injected to the Internet and

now there are more than 120 homogeneous and 170

heterogeneous traffic types passing through the

Internet (Dally, 2001). The most common

communication scenario in today Internet begins

with a client requesting a server for some data. After

receiving the request, the server prepares the

requested data and sends it to the client. This is the

simplest and of course the most common practice in

data communication over the Internet and there are

more than 1 billion of such data transactions per

second in today Internet (Dally, 2001). Among all

different traffic types passing through a general

purpose network (not necessarily the Internet) there

are some special traffic types which have some

constraints (often time limits). A very good example

for this type of traffic is audio and video streams. In

addition to these data streams which can be lossy,

there are some other data streams which have to be

lossless and no packet can be dropped during routing

(Ni, 1993), (Ghiu, 2000), (Shin, 1996), (Kermani,

1979), and (Glass, 1992). For these traffic types

some metrics are defined which are known as QoS

(Quality of Service) constraints. A long list can be

written for QoS constraints since the most common

constraints are delay and jitter.

Now we are faced with a new problem where

some packets belong to a data stream which has

some QoS constraints while the constraints should be

met. A very simple solution to the problem is to

dedicate a well qualified route for these traffic types.

While it is very expensive and impractical, each QoS

constraint raises new requirements. A better and

more reasonable solution is to establish a relatively

well qualified and general purpose network and to

use this single network for communication of all

traffic types. This is what happens in many general

purpose networks including the Internet.

To meet QoS constraints using a general purpose

network, there should be some ways to drop or

postpone the packets having no constraints or are of

lower service quality. The common practice towards

this goal is to use priority queues. Although this

solution can meet many QoS constraints, it does not

utilize network resources (e.g. bandwidth) optimally.

This is due the fact that each router routes

independently from other routers (non-neighboring

routers) and also does not take into account the

609

Mottaghi M. and Daneshtalab M..

USING ROUTING AGENTS FOR IMPROVING THE QUALITY OF SERVICE IN GENERAL PURPOSE NETWORKS.

DOI: 10.5220/0003434906090613

In Proceedings of the 1st International Conference on Pervasive and Embedded Computing and Communication Systems (SAAES-2011), pages

609-613

ISBN: 978-989-8425-48-5

 2011 SCITEPRESS (Science and Technology Publications, Lda.)



Figure 1: The proposed structure.

entire network capacity. In fact, in the conventional

routing method, each router routes individually

regardless of what other routers are doing and where

in the network some extra bandwidth exists.

In this paper, we have proposed a new routing

method which using context-aware routing agents to

switch between minimal and non-minimal routes.

This enables each router (agent) to alter the routes

such that all network capacity is utilized. This is in

contrast to the conventional technique in which even

if routers are context-aware, the entire network

capacity cannot be utilized.

2 THE PROPOSED ROUTING

TECHNIQUE

The proposed routing technique attempts to utilize

all network resources such that the packet loss

becomes less and higher quality services become

feasible.

The drastic difference between the proposed and

the conventional routing techniques is that in the

proposed technique each router uses both minimal

and non-minimal paths based on the entire network

information and other routers' experience. In

contrast, the conventional routers just use minimal

paths to route packets in the network.

2.1 Context Awareness

If we know the attributes of the traffic (like inter-

arrival times, acceptable delay, etc.) we can decide

whether or not a dedicated route is needed. When a

new traffic type is injected to the network the

receiving agent tries to recognize its type. It

measures some attributes of the incoming data

stream. If it could not recognize the traffic type it

asks a neighboring agent (router) to recognize the

traffic type. If agents could not recognize the traffic

type it is considered as a new type and the receiving

agent (at one of network edges) is responsible for

learning how to treat it. The learning process is

based on reinforcement learning methods (Sutton,

2002). When a new traffic type comes to the

network the routing system does not perform

necessarily well and it takes some time for an agent

to learn how to treat this type of traffic. The learning

process is even not always convergent and some

controlling criteria should be applied to avoid the

system from oscillation. These criteria (heuristics)

are often not mathematically proven; we used the

number of hops that a packet visits as the criterion.

2.2 Emergence

After the traffic type is recognized the expert agent

is found too. The expert agent is the agent who

knows how to treat the coming traffic. It evaluates

the entire network capacity and determines whether

PECCS 2011 - International Conference on Pervasive and Embedded Computing and Communication Systems

610

or not the existing routes can accommodate the

newly coming traffic. If no accommodation problem

exists then the new traffic flows through the network

without any modification in the network. If any

shortcoming was found, the expert agent looks for

agents who can provide more bandwidth. If it was

successful in finding any helping agent, the two

agents issue reconfiguration commands to

reconfigure the network such that new QoS

constraints are met. Otherwise newly coming packet

are all dropped, since the network can not serve

them.

If the expert agent realizes that a reconfiguration

in the network is required and if it is successful in

convincing other agents to reconfigure the network,

a new route emerges which can accommodate the

newly coming traffic.

2.3 Collaboration

The whole routing in the proposed technique is

collaborative; collaboration among agents can be

seen in three different situations. First, when a traffic

type is to be recognized, agents send recognition

messages to their neighbors. Second, while looking

for more bandwidth the expert agent requests its

neighbors to provide it with more bandwidth.

Finally, when the expert and the helping agent want

to reconfigure the network they issue

reconfiguration commands to their neighbors.

2.4 Router's Hardware Core

In this section, implementation details of the cited

router are described. As shown in figure 1, each

input port includes an input buffer and a controller

which are used for temporary storage of packets.

The controller of each input port is responsible for

two tasks:

1. Receiving packets from the link and requesting

the crossbar arbiter for packet injection grant.

Also, controlling the buffer status including

empty and full states.

2. Calculating the sign of the rate at which the

buffer is getting full: positive rate indicates that

the buffer is getting full and negative rate

indicates that the buffer is getting empty. This

sign is compared with the buffer status to

activate congestion signal. Each input port has a

congestion signal through which it informs its

adjacent router whether it is congested or not.

Therefore, the owner of an input port may be

considered a hotspot for its adjacent router and

the adjacent router tries to remove this problem.

Another part of this router is the crossbar which

establishes a connection from an input port to an

output port. Since a crossbar can serve only a single

port at a time, it uses an arbitration for the crossbar

to arbiter among the input ports which are not

empty. Afterward, the crossbar is dedicated to the

granted input port.

Figure 2: Address decoder module.

Address decoder is another part of the router taking

the header of the packet stored in the input buffer. It

determines the routes according to both minimal and

non-minimal path adaptive routing algorithms. The

routing algorithm is based on HAMUM

(Daneshtalab, 2011) which is an adaptive minimal

and non-minimal routing algorithm. Then, the

address decoder selects a route which is less

congested, i.e. if the route resulted from the minimal

path adaptive routing algorithm is congested, the

router does not take this route and instead it uses the

result of the non-minimal path adaptive algorithm.

In order to know if the input buffer status is

congested or not, we should measure the rate at

which the data come into the buffer. To do so the

simple module depicted in Figure 3 is used. As

shown in Figure 3, since routers are locally

synchronous and globally asynchronous, no timing

source is used for reading from and writing to the

buffer. Therefore, to obtain the rate at which the

buffer is getting full, the number of filled buffer

cells at each rising edge of router's internal clock

new

) is compared with that of the previous rising

edge (N

old

), i.e. N

new

old

shows that the buffer is

getting full while N

new

old

indicates that the buffer

is getting empty.

Buffers of this router have two special signals

W_Empty and W_Full. When the number of empty

cells of the buffer is smaller than a threshold value

(75%), W_Full (Warning Full) is activated warning

that most buffer cells are full. Congestion status,

which utilizes the W_Full signal and the buffer

filling rate, is used to inform adjacent routers

regarding the input port congestion condition. When

a packet reaches the input port waits until other

previously arrived packets leave the input buffer.

USING ROUTING AGENTS FOR IMPROVING THE QUALITY OF SERVICE IN GENERAL PURPOSE NETWORKS

611

When it is the packet's turn to be routed and the

crossbar arbiter allows the packet to be sent to one

of the output ports, packet header is delivered to the

address decoder to find an efficient non-congested

route to the destination. Thereby, both minimal and

non-minimal path adaptive routing algorithms are

simultaneously loaded to specify two minimal and

non-minimal routes. Afterward, the result of

minimal path adaptive algorithm of HAMUM

checks the corresponding congestion signal coming

from the neighbor input port. If the input port of the

corresponding neighbor router is not congested the

packet is sent to that output port. Otherwise, the non-

minimal route is selected.

Figure 3: Congestion detection module.

3 SIMULATION RESULTS

To evaluate the proposed scheme and to compare it

with existing routing schemes, we have simulated

our proposed scheme with the network simulator

(NS-2, 2006). The agents were implemented in C

(.m files) and simulated using MATLAB (Matlab,

2006). The network core in our simulation has 150

routers through which 80 different traffic types were

flowing (a total of 150000000 packets per second).

For traffic generation we have used NS packages

developed by Ulrich Fiedler for http traffic type

(http traffic generator, 2006) and video streams

(TES video traffic generator, 2006). For other traffic

categories we used NS package BonnTraffic

(scenario based traffic generator, 2006) which is a

scenario based traffic generator. We also generated a

traffic pattern with high QoS constraints, using

MATLAB, and injected to the network.

Results are depicted in Figure 4 to Figure 6. As

shown in Figure 4 the proposed technique has less

delay in comparison with that of priority queues, but

as compared to FIFO it has longer delay. This is

because in FIFO there are more packet loss than in

the proposed technique and these packets (since

dropped) are not taken into account in delay

calculation. In Figure 5 the proposed routing

technique is compared with other conventional

routing techniques in terms of network utilization.

As expected, the proposed technique has better

network utilization. The best network utilization of

the conventional techniques belongs to WFQ

(weighted fair queuing) which is only 30%; network

utilization of our technique is 72% which is by 42%

better than WFQ. This is because in the conventional

techniques no route is altered for QoS purposes,

whereas in the proposed technique the routers can

temporarily form new highways (collections of

routes) by changing the direction of some hop-to-

hop routes. In fact in the proposed technique new

routes emerge.

Figure 4: Delay and Jitter comparison among 4 routing

techniques: FIFO, Priority Queues, WFQ and the proposed

technique.

Figure 5: Network utilization of 4 techniques: FIFO,

Priority Queues, WFQ and the proposed technique.

As depicted in Figure 6, the proposed techniques

cause fewer packet losses than that of FIFO, Priority

queues and WFQ. This emanates from context-

PECCS 2011 - International Conference on Pervasive and Embedded Computing and Communication Systems

612

awareness and collaboration among agents. In the

conventional routing techniques when QoS

constraints cannot be met, the packets are simply

dropped whereas in the proposed technique, agents

collaboratively try to modify the routes such that

new QoS constraints can be met. The efficiency of

the context-awareness along with the collaboration

of agents result in the emergence of new efficient

routes helping the entire routing system drop fewer

packets. Despite having longer average delay, the

proposed routing technique has a higher throughput

and has lower packet loss (Figure 6).

4 CONCLUSIONS

We have proposed a new agent-based routing

method for general purpose networks. Each router

decide to route the packets to either the minimal path

or non-minimal path based on the presented learning

technique among agents (routers). All agents are

aware from their colleagues' status and when a new

traffic type enters the network all agents try to

collaboratively detect its type and route packets

according to their past experience. Simulation results

show that the proposed model can improve QoS of

the network by 12%.

Figure 6: Throughput and Packet Loss of our technique in

comparison with those of FIFO, Priority Queues and

WFQ.

REFERENCES

Dally, W. J., Towles. B., 2001. Internet traffic diversity

analysis. In Proc. DAC, pages 684.689.

Ni. L. M., McKinley, P. K., 1993. A survey of wormhole

routing techniques in direct networks. In IEEE Tran.

on Computers, 26:62.76.

Chiu, G., 2000. The Odd-Even Turn Model for Adaptive

Routing. In IEEE Tran. On Parallel and Distributed

System, Vol.11, NO. 7.

Shin, K. G., Daniel, S. W., 1996. Analysis and

implementation of hybrid switching. In IEEE Tran. on

Computers, 45(6):684.692.

Kermani, P. Kleinrock, L., 1979. Virtual cut-through: a

new computer communication switching technique. In

Computer Networks, volume 3, pages 267.286.

Glass, C. J., Ni, L. M., 1992. The Turn Model for

Adaptive Routing. In Proc. Symp. Computer

Architecture, pp. 278-287.

NS-2, network simulator 2, retrieved on 2006-06-25,

http://nsnam.isi.edu/nsnam/index.php/

Matlab, 2006. http://www.mathworks.com/

products/matlab/.

Http traffic generator, NS-2 package, retrieved on 2006-

06-22, http://www.tik.ee.ethz.ch/~fiedler/provisioning.

TES video traffic generator - NS-2 package, retrieved on

2006-06-22, http://nsnam.isi.edu/

nsnam/index.php/Contributed_Code

Scenario based traffic generator, NS-2 package, retrieved

on 2006-06-22, http://web.informatik.uni-

bonn.de/IV/BoMoNet/

Sutton, R. S., Barto, A. G., 2002. Reinforcement Learning:

An Introduction. MIT Press, Cambridge.

Blaha, M. R., Premerlani, W., 2004. Object- and Agent-

Oriented Modeling and Design. McGrawHill.

Daneshtalab, M., Ebrahimi, M., Xu, T. C., Liljeberg, P.,

Tenhunen, H., 2011. A Generic Adaptive path-based

routing method for MPSoCs. In Journal of Systems

Architecture (JSA-elsevier), Vol. 57, No. 1, pp. 109-

120.

USING ROUTING AGENTS FOR IMPROVING THE QUALITY OF SERVICE IN GENERAL PURPOSE NETWORKS

613