Simulation and Results

Our simulation topology is shown in Figure 2.8. Some variables settings, such as link delay and link capacity, are indicated in the figure.

Figure 2.8 Simulation topology1

There are one UDP flow and two TCP flows in our simulation. The UDP flow is from node S0 to node D1 and is attached by a CBR traffic that has 1 Mb sending rate during the time 40 seconds to the time 100 seconds. One of the TCP flows between node S1 and node D1 (denoted TCP1) exists from the time 0 seconds to the time 100 seconds, and the other flow between node S2 and node D2 (denoted TCP2) exists between the time 20 seconds and 100 seconds. Both two TCP flows are FTP. The total simulation time is 100 seconds. The error model of the wireless links, simulated by two-state Markov chain, is turned on at the time 60 seconds and its average error rate is equal to 0.22. More details about this wireless error model will be described in Chap. 4.

A. Effect of different upper bound and lower bound of gray zone

In this simulation, we want to verify the effect of different upper bounds and lower bounds of gray zone on our packet loss classification algorithm. The upper bound and the

lower bound of the gray zone is determined by αand β, as shown in Eq. 2-5 and Eq.

2-6. When β is fixed as 0.1, we change α from 0.99 to 0.7. The sender uses TCP NewReno to control the data sending rate when network is congested.

The required parameters as described in the above sections are listed below:

--- Search window: 16

--- The threshold of decrease trend th_inc: 0.5 --- The threshold of decrease trend thdec: 0.5

The delay trend scheme is used to be compared with our proposed method and its parameters are set as below.

---γ: 1/16

--- The threshold of delay trend th: 0.5

Then we measure the accuracy A of the flow TCP1 and show the results in Table 1.

Table 1 Accuracy A of TCP1, fixed beta

Alpha

Table 2 Accuracy A of TCP1, fixed alpha

Beta

method 0.01 0.05 0.1 0.15 0.2 0.25 0.3

Delay trend 1 1 1 1 1 1 1

Our scheme 0.92 1 1 1 1 1 1

In Table 1, our method shows better accuracies when α is 0.99 or 0.95 and the accuracies form the delay trend scheme and our method get the same results when α is equal to or less than 0.9. Our method gets high accuracy regardless of the very high value of α. When β is larger than 0.05, our scheme and delay trend scheme get stable

accuracies in Table 2. It means that β effects the accuracy of the packet loss classification slightly.

We modify the actions of TCP NewReno in response to the wireless loss and the details about the modification will be described in Chap. 3. The simulation results using the modified TCP NewReno are given in Table 3 and Table 4.

Table 3 Accuracy A of TCP1, fixed beta, modified TCP NewReno

Alpha

method 0.99 0.95 0.9 0.85 0.8 0.75 0.7 Delay trend 0.75 0.79 0.85 0.9174 0.9174 0.9174 0.9174 Our scheme 0.82 0.8875 0.8875 0.9175 0.9175 0.9175 0.9175

Table 4 Accuracy A of TCP1, fixed alpha, modified TCP NewReno

Beta

method 0.01 0.05 0.1 0.15 0.2 0.25 0.3 Delay trend 0.9 0.83 0.85 0.889 0.889 0.889 0.875 Our scheme 0.9 0.8875 0.8875 0.889 0.889 0.889 0.875

From the simulation result, our scheme shows better accuracies than the accuracies gotten from the delay trend scheme.

B. Effect of different classification threshold

In this simulation, we vary the threshold of increase trend th_inc and the threshold of delay trend th from 0.5 to 0.4. According to the simulation results above, we set the parameter α to be 0.8 and the parameter β to be 0.2. Other parameters are the same as above. The modified TCP NewReno is used to control the network congestion.

The simulation results according to different classification thresholds are shown in Table 5 and Table 6.

Table 5 Accuracy comparison, threshold 0.5 th=0.5

th_inc=0.5

Delay trend scheme Our scheme

Ac Aw A Ac Aw A

TCP1 1 0.831 0.907 1 0.813 0.936

TCP2 1 0.813 0.90 1 0.615 0.928

UDP 0.433 0.996 0.980 0.5 0.979 0.965

Table 6 Accuracy comparison, threshold 0.4 th=0.4

thinc=0.4

Delay trend scheme Our scheme

Ac Aw A Ac Aw A

TCP1 1 0.78 0.926 1 0.813 0.936

TCP2 1 0.46 0.899 1 0.615 0.928

UDP 0.433 0.99 0.983 0.533 0.975 0.963 From the Table 5 and 6, our method gives the same accuracies that measured on the flow TCP1 and the flow TCP2. The performance of our method is quite steady at different values of delay trend threshold; in other words, the proposed method is more insensitive to the threshold. However, delay trend scheme shows better accuracies in response to the flow UDP.

C. Simulation on different topology

We use different topology to evaluate our packet loss classification algorithm and the delay trend scheme. The new topology is shown in Figure 2.9. The link delay and link capacity are labeled above the link. The wireless link is between node W8 and node M0.

Figure 2.9 Simulation topology2

There are three TCP flows and all of them are FTP. The traffic is setting as following.

---flow TCP1: from node W0 to node M0, 0~100 seconds ---flow TCP2: from node W1 to node W5, 20~60 seconds ---flow TCP3: from node W4 to node W7, 40~80 seconds

The total simulation time is 100 seconds. The error model of the wireless links is turned on at the time 60 seconds and its average error rate is equal to 0.034.

The required parameters about our method and the delay trend scheme are listed below:

---α=0.8 ---β=0.2

--- Search window: 16

--- The threshold of decrease trend th_inc: 0.5 --- The threshold of decrease trend thdec: 0.5 ---γ: 1/16

--- The threshold of delay trend th: 0.5

We calculate the accuracies of flow TCP1 and show the results in Table 7.

Table 7 Accuracy comparison for topology2

Ac Aw A

Delay trend scheme 0.58 0.8 0.68 Our method 0.75 0.75 0.75

Our packet loss classification algorithm shows better accuracies and consequently our method is better than the delay trend scheme in this situation.

In the next chapter, we illustrate TCP congestion control algorithm and how we modify the congestion control algorithm in response to the wireless loss.

Chapter 3 TCP Congestion Control