Sunteți pe pagina 1din 6

2010 5th International Symposium on Telecommunications (IST'2010)

An Efficient and Class based Active Queue Management for Next Generation Networks
Amir Hossein Mohajerzadeh, Reza Monsefi, Mohammad Hossein Yaghmaee, Nazbanoo Farzaneh
Department of Computer Engineering, Ferdowsi University of Mashhad, Mashhad, Iran ah.mohajerzadeh@stu-mail.um.ac.ir, {monsefi, hyaghmae}@ferdowsi.um.ac.ir, Farzaneh@stu-mail.um.ac.ir
Abstract- Delivering desired QoS to the applications is one of the most important challenges in next generation networks. Designing efficient queue management in midway nodes leads to improvement in NGN efficiency. This paper presents fair Efficient active Queue Management (EQM) mechanism for congestion avoidance and control in next generation networks. Proposed AQM mechanism uses instantaneous queue length to detect incipient congestion. Different traffics with various QoS requirements are routed in NGN. EQM uses linear optimization problem to control congestion in network nodes. Switching equipments drop arrival packets based on optimization problem outcomes. Proposed AQM mechanism is designed to accompany a transport-layer congestion control protocol such as TCP. We have simulated EQM using Opnet. Results show that proposed mechanism achieves its goals1. Keywords-Active Queue Management; Congestion Management; Drop Probability; Fairness; Optimization Problem

TCP traffic. However non-responsive traffic e.g. User Datagram Protocol (UDP) traffic may still cause congestion due to its greedy behavior. Thus, the need arises for hop by hop congestion control mechanisms. Most deployed mechanism employed by routers to manage their queues, is first-come-firstserve mechanism [2]. Nowadays, many different approaches have produced different AQM policies [3, 4]. AQM policies, when properly used, provide better network utilization and lower end-to-end delays than tail drop. This paper presents a new active queue management based on linear optimization problem, which is called fair Efficient active Queue Management (EQM). Following goals have been considered for EQM. Serving traffics based on their QoS requirements which are defined by sources in intermediate network nodes; to stabilize AQM algorithm action for different load conditions, and to minimize end to end delay and packet loss as much as possible. EQM has tiny computational overhead and is easy to implement. Proposed mechanism considers different parameters based on its goals; these parameters are related to traffics importance, delay tolerance and packet loss tolerance respectively. The performance of EQM has been comprehensively evaluated and compared against existing AQM schemes (RED and FIFO) through Opnet [5] simulations. Simulation results show that EQM achieves its goals. This article has following sections. Section 2 reviews the related works. Section 3 describes the proposed AQM in details. In Section 4 simulation results are shown. Finally, in section 5 we conclude the paper. II. Related Works Different AQM mechanisms have been designed for IP based networks since the time internet deployed. Traditional techniques drop incoming packets managing router queue. The simplest mechanism adopted by routers to manage their queues, is FIFO. This policy, known as tail drop, has two main disadvantages [2]. It notifies sources about congestion only when the queue is already full. As a consequence, packets may suffer high delays when they meet long queues. Moreover, sources may

I.

Introduction

Growth in amount of internet traffics leads to increasing attention to congestion control in intermediate nodes. When the amount of incoming traffics to a node is larger than the amount of outgoing traffic, congestion will be occurred. Most of congestion control methods are divided into to main parts: end to end and hop by hop. End to end algorithms are deployed at the end host where the transport protocol is responsible for detecting congestion in the network. Hop by hop algorithms are implemented in the intermediate network routers. Depended on detected congestion level in the network, end to end algorithms adapt application sending rate. This mechanism is, more popularly known, as end to end congestion control employed by transport protocols such as the Transmission Control Protocol (TCP). In hop by hop algorithms, the intermediate network routers are responsible for detecting oncoming as well as subsisting congestion and provide feedback to the sender indicating the network condition [1]. End to end algorithms work well for traffic that is responsive to congestion e.g.
1

- This research has been supported by ITRC under contract T-500-9966

978-1-4244-8185-9/10/$26.00 2010 IEEE

255

experience multiple losses in packet bursts which arrive at full queues. The second drawback of the tail drop policy is that few connections may monopolize the queue space and prevent other connections to get a share of the link bandwidth. In other words, traffics with different priorities should be served based on their priorities. In traditional AQM policies traffics are treated the same. To eliminate the tail drop disadvantages and to anticipate the source answers to incipient congestion situations, AQM policies [2] have been proposed. AQM uses the principle of the feedback of congestion to end hosts by the mechanism of dropping/marking packets at router queues. The end hosts then react to the dropping/marking of packets by reducing their transmission rate. Consequently, the queue length in routers is reduced and the end-to-end delay experienced by packets is also reduced. Moreover, this mechanism ensures a more efficient use of the network resources by reducing the loss of packets occurring when queues overflow. One of the seminal AQM policies is random early detection (RED) [6]. RED adopts the preemptive dropping, or marking, of packets when the average queue length ranges between the minimum and maximum threshold. More precisely, the probability of packet dropping/marking increases linearly between the minimum and maximum thresholds, and equals zero for average queue lengths below the minimum. All packets are dropped if the average queue length increases above the maximum threshold. In the last several years, objections have been raised against RED [7,8,9], among which the difficulty of properly setting its parameters according to network conditions, and the dependence of the queue length in steady state from the number of flows. Some other AQM mechanisms that have been proposed include SRED [10], Blue [11], Proportional Integral (PI) controller [12] and REM [3]. III. Proposed Cost Function As discussed before, considered parameters in the proposed queue management mechanism are packet loss and end-to-end delay. However, to achieve the main objectives of the proposed mechanism, we have used many other parameters beside. They will be explained in section 3.a. We have considered two types of nodes in IP based networks, source node and intermediate node. Source node, generates and sends traffics to destination through network using intermediate nodes. Intermediate nodes are existing routers in middle of paths. Source node model is shown in figure 1, and the intermediate node model is plotted in figure 2. The purpose of the proposed queue management mechanism in intermediate nodes is to deliver the desired QoS requirements to traffics.

1
Traffic Generator

Class 1 Outgoing Class 2

Class n

Figure 1 - Source node model (traffic generator)

Figure 1 models network layer of intermediate nodes. Traffic generator, generates data in application layer and delivers it to network layer. Input data is divided into different classes based on their QoS requirements. Finally, outgoing section puts data in data link layer to be sent. In figure 1 portion of traffic classes 1,2,...,N, is determined by 1 , 2 ,..., n coefficients respectively.

Classifier
P1 P2 Pi
Figure 2 - internal model of network intermediate nodes

Scheduler

Each node consists of a set of queues. Each queue is specifically considered for storing one class of traffic. Queue priority is determined based on traffic class. Higher priority means that intermediate nodes attempt more to satisfy desired QoS for corresponding traffic. The more, we provide resources to a traffic class, better it receives the service. When a packet reaches an intermediate node, at first step it is delivered to classifier. Classifier delivers packets to its corresponding traffic queue based on the quality of service profile of them. IP header fields are used as proof to classify packets. The most important fields are TOS and the destination IP address. After recognizing the traffic class of each packet, intermediate node delivers packet to Pi unit. Pi unit makes the main decision about a packet; it decides on dropping packet or putting it along the queue. Similar to other queue management mechanisms, proposed queue management mechanism uses packet dropping to regulate ingoing traffic. EQM is developed independent from transport layer protocols, but if the sender uses network responsible protocol (such as TCP), total system efficiency considerably increases. Pi unit drops packets based on probability which is determined by the proposed mechanism. As shown in

256

Figure 2, there is a Pi unit corresponding to each queue. Pi unit values is determined using linear optimization which is described in section 3.a. In this article scheduler unit works as priority based round robin; it means that each queue portion in round robin mechanism depends on its priority. In EQM, when a queue has higher priority; it will give a larger portion in the scheduler. a) Problem Formulation

and their own delay and packet loss priorities network designer should determines coefficients i and i as input for optimization problem.

i + i = 1 term

is

considered base on packet loss and delay parameters that are in contrast to each other in AQM. This means the higher the queue delay priority the lower is the packet loss priority in the opposite side. traffic rate for ith queue of the intermediate node (After passing of Pi eliminator unit). As it is shown in Figure 2, traffic passes eliminator unit before entering queue and i is real arrival traffic rate to queue. By applying a negative coefficient to i variable, in optimization process the aim changes to maximize its value. As mentioned earlier all the queues that are considered in the intermediate nodes are type of M/M/1/K. wi is calculated using set of equations, related to queue theory, which are described in details as follows [13]. wi , waiting time, for the M/M/1/K queue is obtained using equation (4):

wi variable is average waiting time and i is arrival

Portion of each traffic class is determined based on 1 , 2 ,..., n coefficients. They have been valued in source node as shown in fig.1. Number of queues needed in the network node is related to the number of traffic classes; for example, if 10 classes of traffic with different priorities exist in the network, 10 independent physical queues in the intermediate nodes are considered. Equation (1) is respected about i values.

1 + 2 + ... + n = 1

(1)

Queues type is M/M/1/K. This model of queue has reasonable similarity to the real world. Corresponding to each queue, there are input and output models. Input model is considered as Poisson. Input traffic passes Pi unit, which is responsible to congestion control, before getting location in the corresponding queue. Intermediate node makes decision about ith traffic packets based on Pi variable. When a packet is delivered to congestion control unit, it will be dropped with probability Pi and will be queued with probability1 Pi . Objective function of optimization problem is shown in equation (2). In equation (3) a set of initial conditions have been presented for optimization problem.

W =

(4)

is

the real arrival rate for queue and L is the

current average of current queue length. M/M/1/K queue has limited capacity. is calculated using equations (5) and (6).

n =

(0 n < K ) 0 (n K )
n

(5)

Considering

as arrival rate for M/M/1/K queues in

min : F = wi i ii
i =1 i =1

(2)

equation (5), by considering limited capacity of M/M/1/K, value can be calculated normally base on the equation (6).

1 + 2 + ... + n= 1 , i + i = 1
Coefficient

i > 1 ,

1 + 2 + ... + n = 1 i

, )

(3

) = * (1 Pk

(6)

determines delay priority of the ith

Pk is the probability of being full for M/M/1/K. But


as mentioned earlier, there is an eliminator ( Pi unit) before each queue in the intermediate node. So finally by considering probability Pi for each queue,

queue and coefficient

determines packet loss

priority of ith queue. If ith traffic requires higher priority of end to end delay, it needs higher value for corresponding i ; and if ith traffic requires higher priority of packet loss, it needs higher value for corresponding i .Coefficients i and i are calculated continually according to the traffic properties and network status. Based on the number of traffic classes

value can be calculated by using equation (7). Pi


is loss rate and 1 Pi is probability of arrival packet entry to ith queue.

= (1 Pk ) * (1 Pi )

(7)

257

Pk value can be calculated using equation (8).


(1 ) k k +1 = 1 Pk 1 k k +1

( 1) ( = 1)

(8)
Figure 3: Topology used in simulation

As value is calculated by using equation (9).

value is calculated using equation (9). value is


determined in equation (4) using equation (7). As follows, using equation (10), L will be calculated.

In figure 3, Left nodes have been considered as source and right nodes as destination. Each source node generates and forwards different classes of traffics based on Poisson processes with different rates. Each node has various forwarding rate in different scenarios. Proposed queuing mechanism has been implemented in intermediate routers. As mentioned in Section 3, EQM manages the queue using an optimization problem. Queue management means to control arrival traffic. There are some basic parameters in equation (2) which determine proposed mechanism efficiency: , , . Network designer should determine the mentioned value according to traffic characteristics. In section 3.a the parameter is announced as queue input traffic rate. is determined in source node based on traffic characteristics. Three types of traffic are used in simulations: low priority, med priority and high priority. Parameter is determined proportional to traffic generation rate in source node. For example, if the source node produces 5000 packet of type of first priority traffic, 3000 packet of type of second priority traffic and 2000 packet of type of third priority traffic, Parameter value for traffic 1, 2 and 3, is determined respectively 0.5, 0.3 and 0.2. Each queues in the optimization process has

=
L=

(9)

(K + 1) k +1
1 k +1

(10)

It must be considered whatever has been mentioned earlier is related to calculating the average waiting

i = * i *(1 Pi )
time for each queue. In fact, queue.

(11)

equations (2-9) is equal to i that is related to the ith

value was used in the

calculation is explained in equation (11).

is packet arrival rate to Pi unit.

Another important section in the intermediate node management is the scheduler. In fact, output rate for each queue should be considered based on its priority. In this paper, output rate for all queues is considered proportional to their input rate. Set of Pi values as a vector is obtained by using the optimization problem at each run. Due to the dynamic nature of network status, considering just one Pi vector for long period of network lifetime is not efficient. Length of Pi vector calculation interval, directly affects the proposed mechanism efficiency. If the interval is selected shorter, Pi vector performance will be more. On the other hand if user increases the interval length, the proposed mechanism performance will be reduced further application. IV. Performance Evaluation In this section, performance of the proposed active queue management mechanism will be evaluated using Opnet simulator [5]. Packet loss rate, queue length and average end to end delay parameters have been used to evaluate AQM proposed mechanism in intermediate nodes. Dumbbell topology which is shown in figure 3 has been used in simulation.

parameters. End-to-end delay and packet loss parameters are two fundamental parameters of proposed queue management mechanism. Parameter specifies the priority level of end-to-end delay parameter and parameter specifies level of priority of packet loss parameters for each queue. Therefore, and parameters have i index (
and is determined based on corresponding queue priority). i index is the identifier of each individual queue in router. As follows the efficiency of the proposed mechanism under various scenarios has been investigated. In figure 4 queue length is displayed.

and

Horizontal axis is time unit and the vertical axis is the wee queue length. In considered scenario, all the sources dont generate data until time 10, and then at time ten all three sources begin to send data. Sources sending rate are respectively 1000, 2000 and 1250 packet per time unit. Parameter for queues 1, 2 and 3, respectively, is 0.5, 0.5 and 0.5. Parameter value for queues 1, 2 and 3, respectively, is 0.6, 0.5

258

P2 shows queue length for priority 2 queue ( =0.5) and P3 shows queue length for priority 3 queue ( =0.4). The influence of and parameters on proposed queue managements efficiency can be realized by comparison results in figure 4. Average queue length for priority 1 queue is equal to 105, for priority 2 queue is 137 and for priority 3 queue is 142. When the parameter value increases for a queue, this means that the importance of end-to-end delay for the corresponding queue is increases too. And in other hand, when the parameter value for a queue has been increased, this means its importance of loss packages has increased.
160 140 120

Delay

and 0.4 values and Parameter values for queues 1, 2 and 3 are 0.4, 0.5 and 0.6 respectively. In figure 4, P1 shows queue length for priority 1 queue ( =0.6),

0.04 0.035 0.03 0.025 0.02

0.015 0.01 0.005 0 9.8 10 10.2 10.4 10.6 10.8 11

P1

P3

P2

Time

Figure 6: end to end delay for all priority queues

In figure 6, the average end-to-end delay for priority 1, priority 2 and priority 3 queue is, 0.022, 0.028 and 0.030 respectively. As it is shown in values, the end to end delay for priority 1 queue has minimum value among the others two queues, because it has maximum value of Parameter. After priority 1 queue, delay value of priority 2 queue with 0.5 for parameter has the lowest value. According what mentioned in section 3, the proposed queue management mechanism, puts arrival packets in queue or drops them, based on Pi values. At following, P 1 , P2 , P 3 values have been shown at

Queue length

100 80 60 40 20 0 9.8 10 10.2 10.4 10.6 10.8 11 11.2

P1

P2

P3

Time

figure 7. P 1 , P2 , P 3 are loss probability for priority 1, priority 2 and priority 3 queue respectively.
1.2

Figure 4. Average queue length for all priority queues


0.7

P1

P2

P3

Dop Probability
P1 P2 P3

0.6

0.8

Average Packet Loss

0.5 0.4 0.3

0.6

0.4

0.2
0.2 0.1 0 9.8 10 10.2 10.4 10.6 10.8 11 11.2

0 9.8 10 10.2 10.4 10.6 10.8 11 11.2

Time

Time

Figure 7: loss probability for all priority queues

Figure 5 : packet loss percent for all priority queues

In figure 5 the average packet loss rates are 0.365, 0.367 and 0.340 for priority 1, priority 2 and priority 3 queue respectively. Data entry rate is equal for all three queues; priority 1 queue should drop higher number of packets to keep queue length shorter.

Parameter value is equal to 0.6, 0.5 and 0.4 for priority 1, priority 2 and priority 3 queue. By assigning lowest value for coefficient , priority 1 queue has higher end to end delay priority than other two queues. Therefore the results of figures 4, 5 are justifiable. Obviously whatever queue length average is less, traffic has less delay. The router must drop more data (see figure 5) in order to keep priority 1 queue length shorter. As follows in figure 6 end to end delay has been investigated for priority 1 (P1), priority 2 (P2) and priority 3 (P3) queue.

Similar to traditional active queue management techniques, in EQM when queue becoming full, drop probability will be 1. It means all packets will be dropped until queue finds empty space (no package will be accepted similar to tail drop). According to figure 7, whenever priority 3 queue length reaches to its maximum length, which is 150, loss probability will be 1. But average loss probability for priority 1 queue is higher than two other queues regardless the times probability is 1. As observable in figure 7, the average loss probability is 0.4, 0.35 and 0.3 for priority 1, priority 2 and priority 3 queue. At following we have considered different scenario by considering new and values. In table 4, influence

of and parameters on proposed queue management mechanism efficiency have been investigated. We have considered wide range for values in table 1 so and influence on proposed mechanism efficiency is clearer.

259

Table 1- Analysis of the

and

parameters influence on
Value of

proposed queue management efficiency


Average delay 0.15 0.28 0.31 Queue length average 70 131 143

value of 0.8 0.5 0.3

0.2 0.5 0.7

time 11 first transmitter increases its sending rate by 24 percent and the performance parameters have been investigated. In figure 13 queue length chart has been drawn. As its shown in figure 10, although at the time 11 input data rate almost increased 0.24 percent, queue length has not changed dramatically for EQM but queue length is increased noticeable for RED. V. Conclusion In this article a new efficient active queue management mechanism, named EQM, has been proposed to control congestion. EQM considers end to end delays and packet loss as goal parameters. Furthermore, it is able to send traffics with difference priorities, based on their QoS requirements. EQM manages intermediate nodes queue using linear optimization. For each queue and parameters determine its priorities based on end to end delay and loss packet. Results of performed simulations show that proposed mechanism has achieved its goals. In future works we will consider scheduling mechanism besides active queue management. Acknowledgment

In figure 8, queue length has been presented for FIFO besides proposed mechanism, when it is used as queue management mechanism. We have used FIFO as scale to show proposed mechanism efficiency. Generation rate for sources, 1, 2 and 3 are equal to 1000, 2000 and 1250, packets per time unit respectively.
160 140 120

Queue length

100 80 60 40 20 0 9.8 10 10.2 10.4

P1 P3

P2 FIFO

Time

10.6

10.8

11

11.2

Figure 8- queue length charts for FIFO mechanism

This Research has been supported by ITRC under contract T-500-9966. References
[1] Van Jacobson, Michael J. Karels, Congestion Avoidance and Control (1988). Proceedings of the Sigcomm '88 Symposium, vol.18(4): pp.314329 [2] B. Braden et al., Recommendations on queue management and congestion avoidance in the Internet, IETF Request Comments, vol. 2309, Apr. 1998. [3] S. Athuraliya, V. Li, S. Low, and K. Yin, REM: Active queue management, IEEE Network, vol. 15, pp. 4853, May 2001. [4] S. Kunniyur and R. Srikant, Analysis and design of an adaptive virtual queue (AVQ) algorithm for active queue management, in Proc. ACM SIGCOMM, San Diego Ca, Aug. 2001, pp. 123134. [5] www.opnet.com [6] S. Floyd and V. Jacobson, Random early detection gateways for congestion avoidance, IEEE/ACM Trans. Networking, vol. 1, pp. 397413, Aug. 1993. [7] M. May, J. Bolot, C. Diot, and B. Lyes, Reasons not to deploy RED, in Proc. IWQoS, Mar. 1999, pp. 260262. [8] W. Fen, D. Kandlur, D. Saha, and K. Shin, A selfconfiguring RED gateway, in Proc. INFOCOM,W. V. Oz and M. Yannakakis, Eds., New York, Mar. 1999. [9] W. Feng, K. Shin, D. Kandlur, and D. Saha, The BLUE active queue management algorithms, IEEE/ACM Trans. Networking, vol. 10, pp. 513528, Aug. 2002. [10] T. J. Ott, T. V. Lakshman, and L. H.Wong, SRED: stabilized RED, in Proc. IEEE INFOCOM, New York, Mar. 1999, pp. 13461355. [11] W. Feng, K. Shin, D. Kandlur, and D. Saha, The BLUE active queue management algorithms, IEEE/ACM Trans. Networking, vol. 10, pp. 513528, Aug. 2002. [12] C.V. Hollot, V. Misra, D. Towsley, Wei-Bo Gong , On designing improved controllers for AQM routers supporting TCP flows, in Proc. IEEE INFOCOM, Anchorage, AK, Apr. 2001, pp. 17261734. [13] Gross, Donald; Carl M. Harris (1998). Fundamentals of Queueing Theory. Wiley.

In figure 9, end to end delay has been presented for FIFO besides proposed mechanism.
0.04 0.035 0.03 0.025

Delay

0.02 0.015 0.01 0.005 0 9.8 10 10.2 10.4 10.6 10.8 11

P1 P2

P3 FIFO

Time

Figure 9- end to end delay chart for FIFO mechanism


140 120

Queue Length

100 80 60 40 20 0 9.8 10 10.2 10.4 10.6 10.8 11 11.2 RED OBQ

Time

Figure 10- queue moment length charts for explosion traffic

Another important factor in AQM efficiency is to keep queue length fixed. As we know transmitters send their data with different rate over time. Queue management mechanism should keep the average queue length fixed despite variant sending rate. One of the advantages of EQM is keeping the average queue length fixed; however transmitter changes its Send rate. To prove proposed protocol efficiency following scenario has been defined. In figure 10 at

260

S-ar putea să vă placă și