Analysis of Packet Loss Bursts

The distribution of the packet losses is also important for the speech quality. The error concealment function typically handles single and double packet losses fairly well, at least if the losses occur during steady-state signals such as vowels and background noise. But for each consecutive packet loss, the error concealment function will attenuate the signal more and more in order to avoid generating severe distortions. If the packet loss bursts are very long, the signal will be completely muted, giving an impression of a lost call for the listener.

The packet loss burst lengths are also important for the performance of the application layer redundancy. Single packet losses can be recovered by 100% redundancy while longer loss bursts either require a larger amount of redundancy or that the distance between the original frame and the redundant frame is increased. Both these solutions would however increase the end-to-end delay.

It is therefore important to also analyze how the packet losses are grouped into packet loss bursts. The users were grouped according to their respective packet loss rate:

• The group with users having less than 2% PLR is probably the most interesting group because it shows the packet loss distribution for the most normal operating conditions when application layer redundancy is not used.

• The group with users having a packet loss rate between 2% and 10% is interesting for the case when single (100%) redundancy is used.

I Users with PLR<=2% I Users with 2%<PLR<=10% Users with 10%<PLR<=20%

Figure 7.37: Histogram of packet loss bursts for the delay scheduler at 55% relative load.

• The group with users having a packet loss rate between 10% and 20% is interesting for the case when more than 100% application layer is used.

The analysis is performed by calculating the occurrences of packet loss burst with burst lengths of one (= single packet loss), two (= double), three (= triple) and up to 15 consecutive packet losses. The histograms of the analysis for the delay scheduler with different system loads are shown in Figures 7.37 to 7.40.

These figures show that for all users, regardless of packet loss rate, most of the packet losses are single losses or doubles. In fact, the vast majority, about 90% of the packet loss occurrences, are either single or double losses. Longer bursts are mainly three or four in a row. There are a few occurrences of long or even very long packet loss bursts, but these are both rare and occur also only for the highest relative load levels.

One concern when developing the HSPA system and the delay scheduler was that there would be a quite big difference in packet loss burst lengths between low and high load levels and also between users with different packet loss rates, i.e. between satisfied and unsatisfied users. The fact that the histograms are fairly similar for all load levels and also fairly similar for different packet loss rates shows that the scheduler manages to distribute the transmission resources quite well among the users, even for the users that are outside the satisfaction criteria.

The concentration of packet losses to singles and doubles means that application layer redundancy should work fairly well, even without the need to use excessive delay between original and redundant frames. Thereby, it seems reasonable to believe that the speech quality

100 90 80 70 60

Figure 7.38: Histogram of packet loss bursts for the delay scheduler at 70% relative load. 100

90 80 70 60

I Users with PLR<=2% I Users with 2%<PLR<=10% Users with 10%<PLR<=20%

Figure 7.39: Histogram of packet loss bursts for the delay scheduler at 90% relative load.

8 50

cd OC

I Users with PLR<=2% i Users with 2%<PLR<=10% i Users with 10%<PLR<=20%

Figure 7.40: Histogram of packet loss bursts for the delay scheduler at 120% relative load.

could be significantly improved by using application layer redundancy for the users that have a packet loss rate that exceeds the satisfaction criteria.

Figures A.12-A.14 show the histograms of packet loss bursts for the max-CQI scheduler at 10% to 55% relative load levels and Figures A.17 and A.18 show the histogram of packet loss burst for the proportional-fair scheduler at 10% to 35% relative load levels. As can be seen in these figures, both the max-CQI and proportional-fair schedulers give a quite large amount of long packet loss bursts, especially as the load is increased to the capacity limit. This is further evidence that these schedulers do not work very well for real-time VoIP traffic.

Figures A.22-A.24 show the histograms of packet loss bursts for the round-robin scheduler from 10% to 80% relative load respectively. As can be seen in these figures, with the round-robin scheduler, the vast majority of the packet loss occurrences are either single, double and triple packet losses, with single losses being the dominant, as long as the system load is lower than the capacity limit. When the capacity limit is reached and exceeded, packet loss bursts with three losses in a row become dominant. This means that the round-robin scheduler is a fairly good choice, at least from a speech quality perspective and at least as long as the system load is kept below the capacity limit.

0 0

Post a comment