Evaluating the utility of media–dependent FEC on VoIP flows Mart´ ın Varela and Gerardo Rubino Irisa – INRIA/Rennes { mvarela,rubino } @irisa.fr QofIS ’04 – Barcelona September 29 th , 2004 • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
1. Outline • Context • Media–dependent FEC • Assessing the perceived quality • Our approach • Some results • Conclusions • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
2. Context We consider the voice quality of a one–way VoIP flow sent over a best–effort IP network. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
Among the factors affecting quality, we find • Encoding parameters • Packet loss recovery mechanisms, such as – error correction – concealment techniques – buffer management • Network parameters, such as – loss rate – loss distribution – end–to–end delay – jitter • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
3. Media–dependent FEC What is Forward Error Correction? It’s the addition of redundant information which can allow to recover data lost in the network. There are two kinds of FEC • media–independent (e.g. Reed Solomon codes, Hamming codes), which work at the bit level • media–dependent (e.g. as implemented in R.A.T.), which take advantage of the media characteristics Media–dependent FEC can be more efficient than media– independent FEC for both bandwidth consumption and delay. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
The idea is to make each audio packet contain a compressed copy of one (or more) previous packets, so as to allow the decoder to recover from the loss of one of those packets. We are interested in quantifying the impact of FEC on perceived quality. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
4. Assessing the perceived quality There are three main ways to evaluate the perceived quality of a VoIP flow, namely Subjective Assessment – a la ITU P ` .800. This kind of test yields the best results, as it is performed by real people. However, it’s very expensive and time consuming, and of course, it can’t be done automatically. Objective Assessment – this kind of evaluation is the one provided by PESQ, PAMS, MNB, etc. It is cheaper and more practical than subjective assessment, but its correlation with subjective results may be low when considering network impair- ments. Pseudo–subjective Assessment – based on Random Neural Networks, this method produces subjective–like assessments at a lower cost, and can be used in real–time. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
In order to perform a Pseudo–subjective Quality Assessment we • identify (at least some of) the a priori quality–affecting parameters, and their relevant values • select a subset of all possible configurations from the resulting pa- rameter space • generate samples distorted by the selected parameter configurations • carry out a subjective assessment for those samples • train and validate a RNN with the results obtained The resulting RNN yields very good quality estimations for any configu- ration in the parameter space considered. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
5. Our approach We present the results of coupling traditional performance evaluation techniques with Pseudo–subjective Quality Assessment in order to quantitatively evaluate the advantages (or disadvantages) of using media–dependent FEC. We focus on voice quality. We considered the following parameters • Loss rate (0%. . . 15%) • Mean loss burst size (1. . . 2.5 packets) • Codec (PCM Linear–16 and GSM) • FEC (ON/OFF) • FEC offset (1. . . 3 packets) • Packetization interval (20, 40 and 80ms) • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
We used a simple network model to analyze network performance. Our model is an M/M/1/H queue, which represents the bottleneck router in the path from the source to the receiver. Although this is a simple model, it allows to get an idea of the network dynamics. Besides, it also allows for comparison with previous work [1, 2] If need be, a more complex model can be used to model the network, or even a simulator or a test-bed. It only needs to be able to provide the parameters needed for the assessment. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
We considered 2 classes of traffic (audio and background). For the background traffic, we started from the following distribution • 50% 40B packets • 25% 512B packets • 24% 1500B packets • 1% 9180B packets which yields an average packet size of about 600B. We used that mean packet size for our model. We also considered that either all audio packets use FEC, or none of them does. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
In order to assess the impact of FEC on the quality, we need to consider three factors, namely • the proportion of flows using FEC • the impact of the added bitrate on network losses • the benefit of FEC in terms of perceived quality • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
Streaming traffic accounts for a very small fraction of the total traffic, around 6% [3]. This means that voice traffic represents an even smaller portion. However, as VoIP applications gain ground, it is reasonable to see an increase on the voice traffic levels. We considered voice traffic fractions ranging from 5% to 50%. Being that voice packets tend to be very small, they decreased the mean packet size. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
6. Some results We observed that while the use of FEC increases the network load, it does not do it very significantly. For instance, adding GSM FEC to a PCM packet, increases its size by about 10%. Therefore, even for a high fraction of voice flows, the added network load remains relatively low. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
We found that the perceived quality evolved in a similar fashion for all the fractions of voice traffic considered. The increase in voice traffic did not increase the loss rate in a noticeable way, although it did slightly increase the mean loss burst size. This, in turn, decreases the efficiency of FEC unless the FEC offset is adjusted. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
7. Conclusions • We found that the addition of FEC increases quality in all the cases studied. • This improvement is consistent for all the load values we considered. • FEC efficiency decreases with load, but it allows for acceptable voice quality even for high load values. • The added traffic does not induce enough losses so that the use of FEC has a negative impact in quality, at least for reasonable values of the load. • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
References [1] Altman, E., Barakat, C., Ramos, V.: On the utility of FEC mecha- nisms for audio applications. Lecture Notes in Computer Science 2156 (2001) [2] Altman, E., Barakat, C., Ramos, V.: Queueing analysis of simple FEC schemes for IP telephony. In: Proceedings of INFOCOM ’01. (2001) 796–804 [3] Fraleigh, C., Moon, S., Lyles, B., Cotton, C., Kahn, M., Moll, D., Rock- ell, R., Seely, T., Diot, C.: Packet-level traffic measurements from the sprint IP backbone. IEEE Network 17 (2003) 6–17 • First • Prev • Next • Last • Go Back • Full Screen • Close • Quit
Recommend
More recommend
