To verify our proposed scheme, we developed a prototype system. In this section, we describe this system and show the experimental results. The ex-periments cover the verification of the accuracy of the per-segment based full passive QoS measurement scheme and the effectiveness of the transmission control based on the per-segment based QoS.
The prototype of Multi-RTCP device and VoIP clients are implemented on Red Hat Linux. The prototype system can monitor 10,000 RTP packets per second and calculate QoS without packet drop and jitter. The VoIP application was implemented on a cellular phone too. Some experiments are performed with the cellular phone.
VoIP client A Multi-RTCP device VoIP client B
Network A Network B
RTP packet stream
Figure 2.16: Configuration of experiments Table 2.2: Configuration of network simulators
Standard deviation of jitter Packet loss
Network A No jitter 0–5%
Network B 0–100 msec 0%
2.6.1 Packet Loss and Jitter
We verify the accuracy of the per-segment based measurement as the first experiment. Figure 2.16 shows the experimental configuration consicts of two networks they have different characteristics. Table 2.2 details configuration of network simulators they simulate each network segment. Network A causes random packet loss of 0–5%, network B causes jitter of 0–100 msec.
Figure 2.17 proves that the per-segment based full passive QoS measure-ment scheme detects the characteristics of each network with a high degree of accuracy. The per-segment based QoS measurement scheme clarifies the degradation of QoS in which network and how much.
2.6.2 Burst Metrics
In the experiments for burst metrics, the Gilbert-Elliott channel model (Fig-ure 2.18) defined in ITU-T Recommendation G.191 [28] was used as the packet loss model. The model is a state machine which has two states, good and bad. The packet loss rates in the good state and the bad state assume that PG = 0.0 and PB = 0.5. The bursty nature of the model is given by a variable γ(0.0≤γ ≤1.0) with a higher value indicating more bursty packet loss. The transition probabilities are set as follows where f is the packet loss rate.
p= (1−γ)(1−(PB−f)/(PB−PG)) (2.10)
0.0 1.0 2.0 3.0 4.0 5.0 6.0
0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0
measured packet loss rate (%)
packet loss rate set up at network simulator B (%) term. A to term. B, network A
term. A to term. B, network B term. B to term. A, network A term. B to term. A, network B
(a) packet loss rate
0 5 10 15 20 25 30
10 20 30 40 50 60 70 80 90 100
measured interarrival jitter (msec)
standard deviation jitter set up at network simulator A (msec) term. A to term. B, network A term. A to term. B, network B term. B to term. A, network A term. B to term. A, network B
(b) interarrival jitter
Figure 2.17: Packet loss rate and interarrival jitter indicated by Multi-RTCP scheme
q= 1−p−γ (2.11)
γ was set to 0.2 for random packet loss, and to 0.8 for burst packet loss.
We employed a network consisting of two segments for the experiments, a combination of random and burst packet loss.
Prior to the experiments, we show a simulation result of burst metrics observed by the end clients without Multi-RTCP scheme in Figure 2.19. γ of each network was set to 0.2, 0.5, and 0.8. The network ofγ = 0.2 andγ = 0.8 are interconnected, and the network ofγ = 0.5 is set as a single network. The simulation result shows that the burst metrics of the interconnected network of γ = 0.2 andγ = 0.8 are similar to that of the single network ofγ = 0.5. It means that end clients cannot distinguish them each other, and Multi-RTCP scheme can be a solution for it.
Separation of packet loss characteristics was performed for verifying the accuracy of the full passive QoS measurement scheme. Figure 2.20 shows the experimental results and comparisons with the simulation result locally generated. Simulation results were obtained with random packet loss, burst packet loss, tandem connection of random and burst packet loss, and their reverse order. The tandem connection model indicates the medium metrics, which can be measured without the full passive measurement scheme. The graphs also include the per-segment based burst metrics measured by the full passive QoS measurement scheme.
The experimental and simulated results are in close agreement; they re-veal that the scheme separates the burstiness of the two different segments exactly. Consequently, the per-segment based full passive QoS measurement
Good Bad p
q 1-q
1-p
Figure 2.18: Gilbert-Elliott model
0 10 20 30 40 50 60 70 80 90 100
0 50 100 150 200 250 300 350 400 450 500
0.0 2.0 4.0 6.0 8.0 10.0
Burst density (%)
Average burst length (msec)
Packet loss rate (%) Burst length, γ=0.8 and γ=0.2 Burst length, γ=0.2 and γ=0.8 Burst length, γ=0.5 Burst density, γ=0.8 and γ=0.2 Burst density, γ=0.2 and γ=0.8 Burst density, γ=0.5
Figure 2.19: Burst metrics observed by the end clients without Multi-RTCP scheme
scheme can handle the network condition of each segment and perform suit-able admission control.
2.6.3 Transmission Control based on Per-segment based QoS Measurement
This section provides the verification result of transmission control consider-ing factors which cause QoS deterioration shown in Section 2.5. While the configuration of the experiments is as shown in Figure 2.16, the configura-tion of network simulators is updated as shown in Table 2.3. Network A is designed with reverse link of 1x EV-DO, which has date rate of 19.2 kbit/s, 38.4 kbit/s, and 76.8 kbit/s, in mind. Although 1x EV-DO switches the
0 10 20 30 40 50 60 70 80 90 100
0.0 1.0 2.0 3.0 4.0 5.0
burst density (%)
packet loss rate (%)
simulation (random) simulation (burst)
simulation (random+burst) simulation (burst+random) experiment (random as 1st segment) experiment (random as 2nd segment) experiment (burst as 1st segment) experiment (burst as 2nd segment)
(a) burst density
0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
0.0 1.0 2.0 3.0 4.0 5.0
gap density (%)
packet loss rate (%)
simulation (random) simulation (burst)
simulation (random+burst) simulation (burst+random) experiment (random as 1st segment) experiment (random as 2nd segment) experiment (burst as 1st segment) experiment (burst as 2nd segment)
(b) gap density
Figure 2.20: Experimental results of burst density
bandwidth of uplink depending on the size of queued data dynamically, it has the option to fix the bandwidth. On another front, network B is designed by simulating forward link of 1x EV-DO, which provides high bit error rates.
The packet loss model is in accordance with the Gilbert-Elliott channel model where PG = 0, PB = 0.5,γ = 0.2.
RTP streams sent by VoIP client A carry sample voices defined in ITU-T recommendation P.50 [29]. The voice is 16 second length and repeated 32 times (Figure 2.21). VoIP client B records the sample voices, then evaluate the sliced voices by ITU-T recommendation P.862, Perceptual Evaluation of Speech Quality (PESQ) [30]. PESQ is a representative of objective voice quality measurement, which ranges from −0.5 (worst) up to 4.5 (best).
Table 2.4 and Figure 2.22 show experimental results in case of fixed trans-mission methods. The shaded columns in Table 2.4 are the best transtrans-mission methods for that configuration. The best methods are selected by PESQ score and bandwidth in Table 2.1.
Figure 2.23 indicates the transition of PESQ in case of fixed transmission methods. The result points that the voice quality dynamically fluctuates depending on the network configuration and the transmission method. Al-though transmission method 1 keeps the voice quality, it is limited by a ceiling of the low bit rate codec. Transmission method 2 degrades voice quality un-der the configuration of 19.2 kbit/s at network A, however the maximum value of PESQ improves because of the higher bit rate codec. Transmission method 3 requires the widest bandwidth because of enabling FEC, then 19.2 kbit/s and 38.4 kbit/s are not enough to keep voice quality. However, when network A provides a bandwidth of 76.8 kbit/s, the method 3 gives the best PESQ score even if packet loss causes in network B. These results indicate that VoIP clients should choose the adequate transmission method depending
on the cause of the QoS deterioration.
For comparison, we show an example of transmission control depending on the end-to-end QoS. We assume that VoIP client begins to send an RTP stream with the method 2, then switch to the method arranged in advance when packet loss is detected. If the congestion is anticipated, VoIP client should switch to method 1 when packet loss is detected. If the assumption is correct, VoIP client can choose the adequate transmission method. However, when bit error is mistaken for congestion, the bandwidth is limited too much as shown in Figure 2.24a. When congestion is mistaken for bit error in an opposite case, enabling FEC further aggravates the congestion as shown in Figure 2.24b.
Figure 2.25 presents transition of PESQ with transmission control con-sidering factors which cause QoS deterioration described in Section 2.5. The VoIP client begins to transmit an RTP stream with the method 2, then switch to method 1 or 3 depending on the QoS information in received extra RTCP RR packets. When the bandwidth of network A is limited to 19.2 kbit/s (Figure 2.25a), the method 2 causes congestion and impairs PESQ score passingly. However, VoIP client detect the congestion in network A around t1, then update transmission method to the method 1. The method requires sufficiently-narrow bandwidth, and improves PESQ score rapidly.
When bandwidth of network A is limited to 38.4 kbit/s (Figure 2.25b), the packet loss rate in network B changes the behavior of transmission control.
When network B causes no packet loss, VoIP client just keeps transmitting RTP stream with the method 2. In case network B causes packet loss, VoIP client tries method 3 around t2 and returns to method 2 at t3, because of insufficient bandwidth. As a result, the proposed scheme improves QoS com-pared with traditional transmission control detailed in Figure 2.24b. Even when bandwidth of network B is expanded up to 78.6 kbit/s (Figure 2.25c), the packet loss rate in network B affects the behavior of transmission control.
When no packet loss is detected in network B, VoIP client keeps transmission method 2. When packet loss is detected in network B, VoIP client moves on to method 3 to enable FEC around t4. The result is an obvious proof of the advantage of the proposed scheme compared with the traditional scheme resulted in Figure 2.24a.