Keywords

1 Introduction

In recent years, the continuous increase of traffic and service types has caused great demand in IP QoS and MPLS (Multi-Protocol Label Switching) DiffServ (Differentiated Service) model became dominant technology in solving IP QoS problems. The technological combination and interoperability [1] of MPLS and DiffServ is to take advantage of explicit routing and fast forwarding of MPLS and the scalability of DiffServ, so choosing MPLS DiffServ environment has more practical meaning. Active Queue Management (AQM) is a mechanism which can drop packets by some strategy before router buffer is full and avoid global synchronization. RFC 2597 recommends that AQM mechanism be used to realize the multiple levels of drop precedence required in the AF PHB. MRED (Multi-level RED) [2] supporting DiffServ and MPLS is an AQM mechanism that can execute different RED policy and calculate drop probability independently for different drop precedence. In this paper, the limitations of MRED are researched and analyzed. Improved schemes are proposed and I-AMRED algorithm is raised. NS-2 platform is used to verify the effectiveness of I-AMRED algorithm. Comparisons of performance parameters such as throughput, packet loss ratio are presented between I-AMRED and MRED in two experiment schemes using NS-2.

2 The Limitations of MRED

MRED maintains multiple sets of RED thresholds [3]. With the continuous development of IP networks some limitations appear.

  1. (1)

    Sensitiveness to RED Parameters Configuration: Configured RED parameters include the maximum threshold (max th ), the minimum threshold (min th ) and the maximum drop probability (max p ). MRED uses these configuration parameters un-self-adaptively. If RED parameters are configured improperly, the fluctuation of average queue length will become violent [4].

  2. (2)

    Instability When Traffic Is Large or Sharply Increased: Essentially, linear relation between drop probability (P b ) and average queue length (avg_q) for each drop precedence leads to the instability of throughput value. When avg_q ranges from min th to max th , P b will increase slowly with the increase of avg_q. When traffic sharply increases, queue buffer would have been filled up, accordingly, throughput will drop and link utilization will reduce [5].

  3. (3)

    Insufficiency to Reflect Congestion Condition: MRED uses avg_q as unique evaluation parameter to reflect congestion change. However, avg_q can’t always reflect congestion condition correctly. When weight value is low and the value of avg_q is high, the condition of congestion may be relieved. At this time, if deciding drop policy only according to avg_q, packet loss ratio will raise.

3 I-AMRED Algorithm

3.1 Basic Ideas of I-AMRED Algorithm

The purpose of I-AMRED is to remove the limitations of RED, and to accordingly reduce packet loss ratio.

  1. (1)

    Import ARED Mechanism: In order to reduce sensitiveness to RED parameters configuration of MRED, import adaptive mechanism of ARED and make avg_q range from min th to max th as far as possible. Two judgment principles is introduced, that is reducing max p value by dividing a factor α when congestion condition of networks is in a low level, and raising max p value by multiplying a factor β when congestion condition of networks is in a high level. Factor α and β are both more than 1, and control RED policy should be more positive or more conservative by monitoring the change of avg_q every certain period.

  2. (2)

    Change Linear Relation Between P b and avg_q: To raise stability of MRED when traffic is large or sharply increased, linear relation between P b and avg_q should be changed to exponential relation as in Fig. 42.1. The purpose of this change is to reduce P b when avg_q approaches min th (queue buffer of router is relatively idle) and to raise P b when avg_q approaches max th (queue buffer of router is almost filled up), accordingly relieving shortage of queue buffer of router before congestion and improving the stability of network throughput.

    Fig. 42.1
    figure 1figure 1

    Exponential relation of drop probability and average queue length in I-AMRED algorithm

    Full size image
  3. (3)

    Import Real-time Queue Length as Another Evaluation of Congestion Condition: To reflect congestion condition more sufficiently and reduce packet loss ratio, import real-time queue length (qlen) as another evaluation criterion of congestion condition of queue buffer of router. Three judgments are added:

    1. (1)

      avg_q ≥ max th : If qlen is less than max th , use qlen instead of avg_q to calculate drop probability, else, drop all arriving packets.

    2. (2)

      qlen ≥ max th : No matter whether avg_q is more than max th , drop all arriving packets.

    3. (3)

      Use avg_q to calculate drop probability except the two conditions of the two judgments above.

When weight value of a certain Behavior Aggregate (BA) is very low and at this time congestion is relieved just now and avg_q is more than max th but qlen has been already less than max th , it is not essential to drop arriving packets and qlen should be used to implement drop policy instead of drop all arriving packet. When qlen is more than max th , queue buffer of router has been filled up, if avg_q is less than max th and calculate P b and let packets access, causing higher packet loss ratio, at this time, all arriving packets should be dropped. Because the three judgments above are added, congestion will be relieved earlier when qlen is more than max th but avg_q is still less than max th . Therefore, packet drop ratio of network will decrease totally.

3.2 Concrete Design of I-AMRED Algorithm

3.2.1 Calculation Equations

I-AMRED is an algorithm which can differentiates calculation from many BAs. In order to look convenient, variables of I-AMRED as below all aim at a certain BA.

$$ avg\_q = (1 - w_{q} ) \times avg\_q\prime + q \times w_{q} $$
(42.1)
$$ P_{b} = max_{p} \times \exp \left( {\frac{{avg\_q - \min_{th} }}{{\max_{th} - \min_{th} }}} \right) - max_{p} $$
(42.2)

In Eq. (42.1), w q is set as weight value of certain level drop precedence, avg_q is average queue length of last time and initial value of avg_q is zero. q is instantaneous queue length of sampling time. When avg_q ranges from min th to max th , P b is calculated as in Eq. (42.2). The relation between P b and avg_q in Eq. (42.2) is exponential as in Fig. 42.1.

3.2.2 Implementation in NS-2

NS-2 simulation platform contain DiffServ module, where MRED is programmed in the files of “dsredq.cc” and “dsredq.h” [6], so I-AMRED is implemented by modifying program code according to Eq. (42.2).

In the file“dsredq.h”, add enumeration variable status and integer variable qlen, and declare function updateIAMREDMaxP().

In the file“dsredq.cc”, define the function updateIAMREDMaxP() in order to update max p parameter periodically according to I-AMRED judgments. Modify the calculation formulae to exponential form and add judgments about qlen into the function enque(). Add the function updateIAMREDMaxP() into the function calcAvg() in order to update max p parameter periodically for every BA.

4 Experiments on NS-2 Platform

4.1 Experiment 1

The network topology of experiment 1 is set as a simple network. Four source nodes send data to four destination nodes through a bottleneck link which is from a core router to an edge router. Three TCP agents are set on three source nodes which are marked with different drop precedence. An UDP agent is set on one source node, in which traffic is the sum of the three TCP sources. Links between edge routers and core router adopt TSW3CM. All nodes are configured as MPLS nodes.

Two schemes are set in experiment 1. First, let four source nodes send data in a high constant rate at the beginning of simulation time. Second, let two source nodes send data at the beginning of simulation time, and then at the 3rd second let other two source nodes send data. The purpose is to verify stability when traffic sharply increases.

4.2 Experiment 2

The network topology of experiment 2 is set as Fig. 42.2, and the purpose is to examine whether I-AMRED will be stable when the network scale is large. C is core router, E1, E2, …, E11 are source edge routers, E0 is destination edge router, S1, S2, …, S11 are source nodes, D is destination node. Every source node has a TCP agent marked by different drop precedence. These TCP agents use random number generator and adopt exponential distribution generating a number as a time interval of TCP transmission. Pareto model is used to generate a random number to assign a size of file needed to be transmitted. The arrival of packet obeys Poisson distribution. The link between C and E0 is bottleneck link. Accordingly, uncertainties of actual network distribution and data transmission can be simulated and stability of I-AMRED can be verified more practical.

Fig. 42.2
figure 2figure 2

Network topology of experiment 2

Full size image

5 Results and Discussions

5.1 Experiment 1

When four source nodes send data in constant rate at the same time, set simulation time as 10 s, and throughput comparison figure between MRED and I-AMRED is shown in Fig. 42.3. So I-AMRED improves throughput on the whole, and is more stable than MRED at the beginning of simulation time. Packet loss ratio comparison between MRED and I-AMRED is shown in Table 42.1.

Fig. 42.3
figure 3figure 3

Throughput comparison when sending rate is constant in experiment 1

Full size image
Table 42.1 Drop packet ratio comparison when sending rate is constant
Full size table

When letting two TCP sources send data at the beginning of simulation time, and then at the 3rd second let another TCP source and UDP source send data at the second scheme in experiment 1, throughput comparison figure between MRED and I-AMRED is shown in Fig. 42.4. So it is clear that when traffic is sharply increased, throughput using I-AMRED increases as simulation time goes on.

Fig. 42.4
figure 4figure 4

Throughput comparison when traffic increases sharply in experiment 1

Full size image

Packet loss ratio comparison between the two algorithms is as Table 42.2, so I-AMRED algorithm make packet loss ratio of TCP (AF PHB) reduce much, but UDP packet loss ratio raises, for drop precedence of UDP packets is high, and traffic of UDP is the sum of three TCP sources. As a whole, I-AMRED make average packet loss ratio decrease by over 10 %.

Table 42.2 Drop packet ratio comparison when traffic sharply increases at 3rd second
Full size table

5.2 Experiment 2

In experiment 2, throughput comparison figure between MRED and I-AMRED is shown in Fig. 42.5, so I-AMRED make throughput even more stable and higher when network topology is more complex. Packet loss ratio comparison between MRED and I-AMRED is shown in Table 42.3.

Fig. 42.5
figure 5figure 5

Throughput comparison of MRED and I-AMRED in experiment 2

Full size image
Table 42.3 Drop packet ratio comparison in experiment 2
Full size table

Four TCP agents selected are examined. It is clear that packet loss ratios of four TCP sources all decrease after using I-AMRED, but the decrease of low drop precedence is more obvious, the decrease of high drop precedence is less. As a whole, average packet loss ratio decreases notably.

6 Conclusion

The results of experiment 1 and 2 indicate that I-AMRED raises throughput especially when traffic is large. And it sharply increases compared with MRED, and notably reduces packet loss ratio of AF PHB. Because of the improvement of throughput and packet loss ratio, sensitiveness to RED control parameter decreases, while stability of network throughput increases. These advantages make I-AMRED algorithm very suitable for current IP networks in which the amount and type of user and traffic expand rapidly. In summary, I-AMRED algorithm is ideally suitable for MPLS networks based on DiffServ.