Intra-flow Loss Recovery and Control for VoIP

Henning Sanneck¹, Nguyen Tuong Long Le¹,
Adam Wolisz², Georg Carle¹

¹GMD Fokus, Kaiserin-Augusta-Allee 31, D-10589 Berlin, Germany
{sanneck, le, carle}@fokus.gmd.de

²Technical University Berlin, Telecommunication Networks Group,
Sekr. FT 5-2, Einsteinufer 25, D-10587 Berlin, Germany
wolisz@ee.tu-berlin.de

Abstract

``Best effort'' packet-switched networks, like the Internet, do not offer a reliable transmission of packets to applications with real-time constraints such as voice. Thus, the loss of packets impairs the application-level utility. For voice this utility impairment is twofold: on one hand, even short bursts of lost packets may decrease significantly the ability of the receiver to conceal the packet loss and the speech signal playout is interrupted. On the other hand, some packets may be particular sensitive to loss as they carry more important information in terms of user perception than other packets.

We first develop an end-to-end model based on loss run-lengths with which we can describe the loss distribution within a flow. These packet-level metrics are then linked to user-level objective speech quality metrics. Using this framework, we find that for low-compressing sample-based codecs (PCM) with loss concealment isolated packet losses can be concealed well, whereas burst losses have a higher perceptual impact. For high-compressing frame-based codecs (G.729) on one hand the impact of loss is amplified through error propagation caused by the decoder filter memories, though on the other hand such coding schemes help to perform loss concealment by extrapolation of decoder state. Contrary to sample-based codecs we show that the concealment performance may ``break'' at transitions within the speech signal however.

We then propose mechanisms which differentiate between packets within a voice data flow to minimize the impact of packet loss. We designate these methods as ``intra-flow'' loss recovery and control. At the end-to-end level, identification of packets sensitive to loss (sender) as well as loss concealment (receiver) takes place. Hop-by-hop support schemes then allow to (statistically) trade the loss of one packet, which is considered more important, against another one of the same flow which is of lower importance. As both packets require the same cost in terms of network transmission, a gain in user perception is obtainable. We show that significant speech quality improvements can be achieved and additional data and delay overhead can be avoided while still maintaining a network service which is virtually identical to best effort in the long term.

Introduction

Considering that a real-time flow may experience some packet loss, the impact of loss may vary significantly dependent on which packets are lost within a flow. In the following we distinguish two reasons for such a variable loss sensitivity:

1. Temporal sensitivity: Loss of ADUs which is correlated in time may lead to disruptions in the service. Note that this effect is further aggravated by some interdependence between ADUs (i.e., that one ADU can only be decoded when a previous ADU before has successfully been received and decoded). For voice, as a single packet contains typically several ADUs (voice frames) this effect is thus more significant than e.g. for video. It translates basically to isolated packet losses versus losses that occur in bursts.
2. Sensitivity due to ADU heterogeneity: Certain ADUs might contain parts of the encoded signal which are more important with regard to user perception than others of the same flow. Let us consider a flow with two frame types of largely different perceptual importance (we assume same size, frequency and no inter-dependence between the frames). Under the loss of 50% of the packets, the perceptual quality varies hugely between the case where the 50% of the frames with high perceptual importance are received and the case where the 50% less important frames are received.

Figure 1: Schematic utility functions dependent on the loss of more (+1) and less (-1) important packets

Network support for real-time multimedia flows can on one hand aim at offering a lossless service, which, however, to be implemented within a packet-switched network, will be costly for the network provider and thus for the user. On the other hand, within a lossy service, the above sensitivity constraints must be taken into account. It is our strong belief that this needs to be done in a generic way, i.e., no application-specific knowledge (about particular coding schemes e.g.) should be necessary within the network and, vice versa, no knowledge about network specifics should be necessary within an application. Let us now consider the case that 50% of packets of a flow are identified as more important (designated by ``+1'') or less important (``-1'') due to any of the above sensitivity constraints. Figure 1 a) shows a generic utility function describing the application-level Quality of Service (QoS) dependent on the percentage of packets lost. For real-time multimedia traffic, such utility should correspond to perceived video/voice quality. If the relative importance of the packets is not known by the transmission system, the loss rates for the +1 and -1 packets are equal. Due to the over-proportional sensitivity of the +1 packets to loss as well as the dependence of the end-to-end loss recovery performance on the +1 packets, the utility function is decreasing significantly in a non-linear way (approximated in the figure by piece-wise linear functions) with an increasing loss rate. Figure 1 b) presents the case where all +1 packets are protected at the expense of -1 packets. The decay of the utility function (for loss rates < 50%) is reduced, because the +1 packets are protected and the end-to-end loss recovery can thus operate properly over a wider range of loss rates indicated by the shaded area. This results in a graceful degradation of the application's utility. Note that the higher the non-linearity of the utility contribution of the +1 packets is (deviation from the dotted curve in Fig.  1 a), the higher is the potential gain in utility when the protection for +1 packets is enabled. Results for actual perceived quality as utility for multimedia applications exhibit such a non-linear behavior.

To describe this effect and provide a taxonomy for different QoS enhancement approaches, we introduce a novel terminology: we designate mechanisms which influence QoS parameters between flows (thus decrease the loss rate of one flow at the expense of other flows) as inter-flow QoS. Schemes which, in the presence of loss, differentiate between packets within a flow as demonstrated in Figure 1 above, provide intra-flow QoS enhancement. As QoS mechanisms have to be implemented within the network (hop-by-hop) and/or in the end systems (end-to-end), we have another axis of classification.

The adaptation of the sender's bitrate to the current network congestion state as an intra-flow QoS scheme (loss avoidance, [16]) is difficult to apply to voice. Considering that voice flows have a very low bitrate, the relative cost of transmitting the feedback information is high (when compared e.g. to a video flow). To reduce this cost the feedback interval would need to be increased, then leading to a higher probability of wrong adaptation decisions. The major difficulty, however, is the lack of a codec which is truly scalable in terms of its output bitrate and corresponding perceptual quality. Currently standardized voice codecs ([17]) usually only have a fixed output bitrate. While it has been proposed to switch between voice codecs ([2]), the MOS (subjective quality) values for the codecs employed do not differ much: e.g., the ITU codecs G.723.1, G.729, G.728, G.726 and G.711 cover a bitrate range from 5.3 kbit/s to 64 kbit/s while the subjective quality differs by less than 0.25 on a 1-to-5 MOS scale ([4], 1: bad, 5: excellent quality). So when the availability of sufficient computing power is assumed, the lowest bitrate codec can be chosen permanently without actually decreasing the perceptual quality.

For loss recovery on an end-to-end basis, due to the real-time delay constraints, open-loop schemes like Forward Error Correction (FEC) have been proposed ([2]). While such schemes are attractive because they can be used on the Internet today, they also have several drawbacks. The amount of redundant information needs to be adaptive to avoid taking bandwidth away from other flows. This adaptation is crucial especially when the fraction of traffic using redundancy schemes is large ([8]). If the redundancy is a source coding itself, like it has often been proposed ([7]), the comments from above on adaptation also apply. Using redundancy has also implications to the playout delay adaptation ([10]) employed to de-jitter the packets at the receiver. Note that the presented types of loss sensitivity also apply to applications which are enhanced by end-to-end loss recovery mechanisms. End-to-end mechanisms can reduce and shift such sensitivities but cannot come close to eliminate them.

Therefore in this work we assume that the lowest possible bitrate which provides the desired quality is chosen. Neither feedback/adaptation nor redundancy is used, however, at the end-to-end level, identification/marking of packets sensitive to loss (sender) as well as loss concealment (receiver) takes place. Hop-by-hop support schemes then allow trading the loss of one packet, which is considered more important, against another one of the same flow which is of lower importance. We employ actual codecs and measure their utility in the presence of packet loss using objective speech quality measurement.

The paper is structured as follows: Section Metrics introduces packet- and user-level metrics. We employ these metrics to describe the sensitivity of VoIP traffic to packet loss in section VoIP Loss Sensitivity. Section Queue Management for Intra-flow Loss Control briefly introduces a queue management algorithm which can be used for intra-flow loss control. In section Intra-flow Loss Recovery and Control, we present results documenting the performance of the proposed mechanisms at both the end-to-end and hop-by-hop level. Section Conclusions concludes the paper.

Metrics

Packet-level metrics

A general Markov model ([19,6]) which describes the loss process is defined as follows:

Let $P( \; l(s) \; \vert \; l(s-m),\dots,l(s-2),l(s-1) \; )$ be the state transition probability of a general Markov model of order $m$, where $l(s)$ is the loss indicator function for the packet with the sequence number $s$. All combinations for the values ($0$ and $1$) of the sequence $l(s-m),\dots,l(s-2),l(s-1)$ appear in the state space. As an example $P( \; l(s)=1 \; \vert \; l(s-2),l(s-1) = 01 \; )$ gives the state transition probability when the current packet $s$ is lost, the previous packet $s-1$ has also been lost and packet $s-2$ has not been lost. The number of states of the model is $2^m$. Two state transitions can take place from any of the states. Thus, the number of parameters which have to be computed is $2^{m+1}$. Even for relatively small $m$ this amount of parameters is difficult to be evaluated and compared.

Table 1: State and transition probabilities computed for an end-to-end Internet trace using a general Markov model (third order) by Yajnik et. al. [9]

Table 1 shows some values for the state and transition probabilities for a general Markov model of third order measured end-to-end in the Internet by Yajnik et. al. ([19]). It is interesting to note that for all states with $l(s-1)=0$ the probability for the next packet not to be lost ($l(s)=0$) is generally very high ($>0.8$, in bold typeface) whereas when $l(s-1)=1$ the state transition probabilities to that event cover the range of $0.15$ to $0.61$. That means that past ``no loss'' events do not affect the loss process as much as past loss events. Intuitively this seems to make sense, because a successfully arriving packet can be seen as a indicator for congestion relief. Andren et. al. ([1]) as well as Yajnik et. al. ([20]) both confirmed this by measuring the cross correlation of the loss- and no-loss-run-lengths. They came to the result that such correlation is very weak. This implies that patterns of short loss bursts interspersed with short periods of successful packet arrivals occur rarely (note in this context that in Table 1 the pattern $101$ has by far the lowest state probability).

Thus, in the following we employ a model ([12]) which only considers the past loss events for the state transition probability. The number of states of the model can be reduced from $2^m$ to $m+1$. This means that we only consider the state transition probability $P( \; l(s) \; \vert \; l(s-k),\dots,l(s-1) \; )$ with $l(s-k+i) = 1 \; \forall \; i \in [0, k-1]$, where $k$ ($0 < k \le m$) is a variable parameter. We define a loss run length $k$ for a sequence of $k$ consecutively lost packets detected at $s_j$ ($s_j > k > 0$) with $l(s_j-k-1) = 0, l(s_j)=0$ and $l(s_j-k+i) = 1 \; \forall \; i \in [0, k-1]$ , $j$ being the $j$-th ``burst loss event''. Note that the parameters of the model become independent of the sequence number $s$ and can now rather be described by the occurence $o_k$ of a loss run length $k$.

We define the random variable $X$ as follows: $X=0$: ``no packet lost'', $X=k$ ($0<k<m$): ``exactly $k$ consecutive packets lost'', and $X\ge k$ ($0<k<m$): ``at least $k$ consecutive packets lost''. With this definition, we establish a loss run-length model with a finite number of states ($m+1$) which gives loss probabilities dependent on the burst length. In the model, for every additional lost packet which adds to the length of a loss burst a state transition takes place. If a packet is successfully received, the state returns to $X=0$. Thus, the state probability of the system for $0<k<m$ is $P(X\ge k)$ . Due to the limited memory of the system, the last state $X=m$ is just defined as ``$m$ consecutive packets lost'', with $P(X=m)$ being the state probability. Given the case of a finite number of packets $a$ for a flow, which experiences $d=\sum_{k=1}^{\infty}k o_k$ packet drops, we have the relative frequency $p_{L,k} = \frac{o_k}{a}$ ($P(X=k)$ for $a \to \infty$) for the occurence of a loss burst of length $k$. An approximation for the expectation of the random variable $X$ can be computed as $p_L = \sum_{k=1}^{\infty}k p_{L,k}$ and identified with the ``mean loss rate''. We can also approximate the state probabilities of the model by the cumulative loss rate $p_{L, cum}(k) = \sum_{n=k}^{\infty}p_{L,n}$ ($0<k<m$) and $p_{L,m} = \sum_{n=m}^{\infty}\frac{(n-m+1)o_n}{a}$ ($k=m$). The transition probabilities for $1<k<m$ can be computed easily as:

$$\begin{eqnarray*} p_{(k-1)(k)} & = & \scriptstyle P(X \ge k \vert X \ge k-1) = \... ...(X\ge k-1)}\ & = & \textstyle \frac{P(X \ge k)}{P(X \ge k-1)} \end{eqnarray*}$$

These conditional loss probabilities again can be approximated by $p_{L, cond}(k) = \frac{p_{L, cum}(k)}{p_{L, cum}(k-1)} = \frac{\sum_{n=k}^{\infty}o_n}{\sum_{n=k-1}^{\infty}o_n}$. For $P(X=m \vert X=m)$ we have:

$$\textstyle p_{L, cond}(m) = \textstyle\frac{\sum_{n=m}^{\infty}(n-m)o_n}{\sum_{n=m}^{\infty}n o_n}$$ (1)

Fig. 2 shows the Markov chain for the model.

Figure 2: Loss run-length model with $m+1$ states

Additionally, we also define a random variable $Y$ which describes the distribution of burst loss lengths with respect to the burst loss events $j$ (and not to packet events like in the definition of $X$ ). We have the burst loss length rate $g_k = \frac{o_k}{\sum_{n=1}^{\infty}o_n}$ as an estimate for $P(Y=k)$. Thus, the mean burst loss length is $g = \frac{d}{\sum_{k=1}^{\infty}o_k} = \frac{\sum_{k=1}^{\infty}k o_k}{\sum_{k=1}^{\infty}o_k} = \sum_{k=1}^{\infty}k g_k$ corresponding to $E[Y]$ (average loss gap, [5]).

The run-length model implies a geometric distribution for residing in state $X=m$. For the probability of a burst loss length of $k$ packets we thus can compute estimates for performance parameters in a higher order model representation (note that here $Y$ represent the random variable used in the higher-order models). For a three state model we have e.g. for $P(Y=k)$:

$$\textstyle \hat{P}(Y=k) = \left\{ \begin{array}{r@{:}l} P(... ...^{k-2} \; (1-p_{22})& \quad k \ge 2 \quad \end{array} \right.$$ (2)

For the special case of a system with a memory of only the previous packet ($m=1$), we can use the runlength distribution for a simple computation of the parameters of the commonly-used Gilbert model (Fig. 3) to characterize the loss process ($X$ being the associated random variable with $X=0$: ``no packet lost'', $X=1$ ``a packet lost''). Then the probability of being in state $m$ can be seen as the unconditional loss probability ulp and approximated by the mean loss ($p_L = p_{L,m}$). Only one conditional loss probability clp for the transition $1 \rightarrow 1$ exists:

$$p_{L,cond} = \frac{\sum_{n=1}^{\infty}(n-1)o_n}{d}$$ (3)

If losses of one flow are correlated (i.e., the loss probability of an arriving packet is influenced by the contribution to the state of the queue by a previous packet of the same flow and/or both the previous and the current packet see bursty arrivals of other traffic, [15]) we have $p_{01} \le clp$ and thus $ulp \le clp$. For $p_{01} = clp$ the Gilbert model is equivalent to a 1-state (Bernouilli) model with $ulp = clp$ (no loss correlation).

Figure 3: Loss run-length model with two states (Gilbert model)

As in equation 2 we can compute an estimate for the probability of a burst loss length of $k$ packets for a higher order model representation:

$$\textstyle \hat{P}(Y=k) = clp^{k-1}(1-clp)$$ (4)

Objective Speech Quality Metrics

Unlike the conventional methods like the Signal-to-Noise Ratio ($SNR$), novel objective quality measures attempt to estimate the subjective quality as closely as possible by modeling the human auditory system. In our evaluation we use two objective quality measures: the Enhanced Modified Bark Spectral Distortion (EMBSD, [21]) and the Measuring Normalizing Blocks (MNB) described in the Appendix II of the ITU-T Recommendation P.861 ([18]). These two objective quality measures are reported to have a very high correlation with subjective tests, their relation to the range of subjective test result values (MOS) is close to being linear and they are recommended as being suitable for the evaluation of speech degraded by transmission errors in real network environments such as bit errors and frame erasures ([18,21]). Both metrics are distance measures, i.e., a result value of 0 implies perfect quality, the larger the value, the worse the speech quality is (in Fig. 5 we show an axis with approximate corresponding MOS values). For all simulations in this paper we employed both schemes. As they yielded very similar results (though MNB results generally exhibited less variability) we only present EMBSD results.

VoIP Loss Sensitivity

Figure 4: Components of the end-to-end loss recovery/control measurement setup.

Figure 4 shows the components of the measurement setup which we will use to evaluate our approach to combined end-to-end and hop-by-hop loss recovery and control. The shaded boxes show the components in the data path where mechanisms of loss recovery are located. Together with the parameters of the network model (section Packet-level metrics) and the perceptual model we obtain a measurement setup which allows us to map a specific PCM signal input to a speech quality measure. While using a simple end-to-end loss characterization, we generate a large number of loss patterns by using different seeds for the pseudo-random number generator (for the results presented here we used $300$ patterns for each simulated condition for a single speech sample). This procedure takes thus into account that the impact of loss on an input signal may not be homogenous (i.e., a loss burst within one segment of that signal can have a different perceptual impact than a loss burst of the same size within another segment). By averaging the result of the objective quality measure for several loss patterns, we have a reliable indication for the performance of the codec operating under a certain network loss condition. We employed a Gilbert model (Fig. 3) as the network model for the simulations, as we have found that here higher order models do not provide much additional information.

Temporal sensitivity

PCM

We first analyze the behaviour for $\mu$-law PCM flows (64 kbit/s) with and without loss concealment, where the loss concealment repairs isolated losses only (speech stationarity can only be assumed for small time intervals). Results are shown for the AP/C concealment algorithm ([11]). Similar results were obtained with other concealment algorithms. Figure 5 shows the case without loss concealment enabled where Gilbert model parameters are varied. The resulting speech quality is insensitive to the loss distribution parameter ($clp$). The results are even slightly decreasing for an increasing $clp$, pointing to a significant variability of the results. In Figure 6 the results with loss concealment are depicted. When the loss correlation ($clp$) is low, loss concealment provides a significant performance improvement. The relative improvement increases with increasing loss ($ulp$). For higher $clp$ values the cases with and without concealment become increasingly similar and show the same performance at $clp \approx 0.3$. Figures 5 and  6 respectively also contain a curve showing the performance under the assumption of random losses (Bernouilli model, $ulp = clp$). Thus, considering a given $ulp$, a worst case loss pattern of alternating losses ( $l(s \bmod 2) = 1, l([s+1] \bmod 2) = 0$) would enable a considerable performance improvement (with $o_k=0 \forall k>1$: $p_{L,cond}=0$, Eq.  3).

Figure 5: Utility curve for PCM without loss concealment (EMBSD)

Figure 6: Utility curve for PCM with loss concealment (EMBSD)

Figure 7: Variability of the perceptual distortion with loss concealment (EMBSD)

As we found by visual inspection that the distributions of the perceptual distortion values for one loss condition seem to approximately follow a normal distribution we employ mean and standard deviation to describe the statistical variability of the measured values. Figure 7 presents the perceptual distortion as in the previous figure but also give the standard deviation as error bars for the respective loss condition. It shows the increasing variability of the results with increasing loss correlation ($clp$), while the variability does not seem to change much with an increasing amount of loss ($ulp$). On one hand this points to some, though weak, sensitivity with regard to heterogeneity (i.e., it matters which area of the speech signal (voiced/unvoiced) experiences a burst loss). On the other hand it shows, that a large number of different loss patterns is necessary for a certain speech sample when using objective speech quality measurement to assess the impact of loss correlation on user perception.

G.729

Figure 8: Utility curve for the G.729 codec (EMBSD)

G.729 ([17]) uses the Conjugate Structure Algebraic Code Excited Linear Prediction (CS-ACELP) coding scheme and operates at 8 kbit/s. In G.729, a speech frame is 10 ms in duration. For each frame, the G.729 encoder analyzes the input data and extracts the parameters of the Code Excited Linear Prediction (CELP) model such as linear prediction filter coefficients and excitation vectors. When a frame is lost or corrupted, the G.729 decoder uses the parameters of the previous frame to interpolate those of the lost frame. The line spectral pair coefficients (LSP) of the last good frame are repeated and the gain coefficients are taken from the previous frame but they are damped to gradually reduce their impact. When a frame loss occurs, the decoder cannot update its state, resulting in a divergence of encoder and decoder state. Thus, errors are not only introduced during the time period represented by the current frame but also in the following ones. In addition to the impact of the missing codewords, distortion is increased by the missing update of the predictor filter memories for the line spectral pairs and the linear prediction synthesis filter memories. Figure 8 shows that for similar network conditions the output quality of the G.729 is worse than PCM with loss concealment, demonstrating the compression versus quality tradeoff under packet loss. Interestingly the loss correlation ($clp$ parameter) has some impact on the speech quality, however, the effect is weak pointing to a certain robustness of the G.729 codec with regard to the resiliency to consecutive packet losses due to the internal loss concealment. Rosenberg has done a similar experiment ([9]), showing that the difference between the original and the concealed signal with increasing loss burstiness in terms of a mean-squared error is significant, however. This demonstrates the importance of perceptual metrics which are able to include concealment (and not only reconstruction) operations in the quality assessment.

Sensitivity due to ADU heterogeneity

PCM is a memoryless encoding. Therefore the ADU content is only weakly heterogeneous (Figure 7). Thus, in this section we concentrate on the G.729 codec. The experiment we carry out is to measure the resynchronization time of the decoder after $k$ consecutive frames are lost. The G.729 decoder is said to have resynchronized with the encoder when the energy of the error signal falls below one percent of the energy of the decoded signal without frame loss (this is equivalent to a signal-to-noise ratio ($SNR$) threshold of $20 dB$). The error signal energy (and thus the $SNR$) is computed on a per-frame basis. Figure 9 shows the resynchronization time (expressed in the number of frames needed until the threshold is exceeded) plotted against the position of the loss (i.e., the index of the first lost frame) for different values of $k$ . The speech sample is produced by a male speaker where an unvoiced/voiced ($uv$) transition occurs in the eighth frame.

Figure 9: Resynchronization time (in frames) of the G.729 decoder after the loss of $k$ consecutive frames ($k\in [1,4]$) as a function of frame position.

We can see from Figure 9 that the position of a frame loss has a significant influence on the resulting signal degradation, while the degradation is not that sensitive to the length of the frame loss burst $k$. The loss of unvoiced frames seems to have a rather small impact on the signal degradation and the decoder recovers the state information fast thereafter. The loss of voiced frames causes a larger degradation of the speech signal and the decoder needs more time to resynchronize with the sender. However, the loss of voiced frames at an unvoiced/voiced transition leads to a significant degradation of the signal. We have repeated the experiment for different male and female speakers and obtained similar results. Taking into account the used coding scheme, the above phenomenon could be explained as follows: Because voiced sounds have a higher energy than unvoiced sounds, the loss of voiced frames causes a larger signal degradation than the loss of unvoiced frames. However, due to the periodic property of voiced sounds, the decoder can conceal the loss of voiced frames well once it has obtained sufficient information on them. The decoder fails to conceal the loss of voiced frames at an unvoiced/voiced transition because it attempts to conceal the loss of voiced frames using the filter coefficients and the excitation for an unvoiced sound. Moreover, because the G.729 encoder uses a moving average filter to predict the values of the line spectral pairs and only transmits the difference between the real and predicted values, it takes a lot of time for the decoder to resynchronize with the encoder once it has failed to build the appropriate linear prediction filter.

Queue Management for Intra-flow Loss Control

Figure 10: ''Differential'' RED drop probabilities as a function of average queue sizes

While we have highlighted the sensitivity of VoIP traffic to the distribution of loss in the last sections, we now want to briefly introduce a queue management mechanism ([13]) which is able to enforce the relative preferences of the application with regard to loss.

We consider flows with packets marked with ``+1'' and ``-1'' (as described in the introduction) as ``foreground traffic'' (FT) and other (``best effort'') flows as ``background traffic'' (BT). Packet marking, in addition to keeping the desirable property of state aggregration within the network core as proposed by the IETF Differentiated Services architecture, is exploited here to convey the intra-flow requirements of a flow. As it should be avoided to introduce reordering of the packets of a flow in the network we consider mechanisms for the management of a single queue with different priority levels. One approach to realize inter-flow service differentiation using a single queue is RIO ('RED with IN and OUT', [3]). With RIO, two average queue sizes as congestion indicators are computed: one just for the IN (high priority) packets and another for both IN and OUT (low priority) packets. Packets marked as OUT are dropped earlier (in terms of the average queue size) than IN packets. RIO has been designed to decrease the $ulp$ seen by particular flows at the expense of other flows. In this work, however, we want to keep the $ulp$ as given by other parameters while modifying the loss distribution for the foreground traffic (FT). This amounts to trading the loss of a ``+1'' packet against a ``-1'' packet of the same flow (in a statistical sense). Fig.  10 shows the conventional RED drop probability curve ($p_0$ as a function of the average queue size for all arrivals $avg$), which is applied to all unmarked (``$0$'') traffic (background traffic: BT).

The necessary relationship between the drop probabilities for packets marked as "-1" and ``+1'' can be derived as follows (note that this relationship is valid both at the end-to-end level and every individual hop): Let $a = a_0 + a_{+1} + a_{-1}$ be the overall number of emitted packets by an FT flow and $a_x, x \in [-1,0,+1]$ be the number of packets belonging to a certain class (where the ``0'' class corresponds to (unmarked) ``best effort'' traffic). Then, with $a_{+1} = a_{-1} = a_{\vert 1\vert}$ and considering that the resulting service has to be best effort in the long term, we have:

$$\textstyle a_0 p_ 0 + a_{+1} p_{+1} + a_{-1} p_{-1}$$ $\stackrel{!}{=}$$$a p_0$$ 
$$a_{\vert 1\vert} (p_{+1} + p_{-1})$$ $=$ $$(a - a_0) p_0$$  
$$p_{-1}$$ $=$ $$2 p_0 - p_{+1}$$ (5)

Due to this relationship between the drop probability for ``+1'' and ``-1'' packets, we designate this queue management algorithm as ``Differential'' RED (DiffRED). Figure 10 shows the corresponding drop probability curves. Due to the condition of $a_{+1} = a_{-1} = a_{\vert 1\vert}$ in addition to the conventional RED behaviour, the DiffRED implementation should also monitor the $+1$ and $-1$ arrival processes. If the ratio of $+1$ to $-1$ packets at a gateway is not $1$ (either due to misbehaving flows or a significant number of flows which have already experienced loss at earlier hops) the $-1$ loss probability is decreased and the $+1$ probability is increased at the same time thus degrading the service for all users. The shaded areas above and below the p$_0(avg)$ curve (Fig. 10) show the operating area when this correction is added.

In [13] it has been shown that using only the conventional RED average queue size $avg$ for DiffRED operation is not sufficient. This is due to the potentially missing correlation of the computed $avg$ value between consecutive $+1$ and $-1$ arrivals, especially when the share of the FT traffic is low. As this might result in a unfair distribution of losses between the FT and BT fractions, a specific $avg_1$ value is computed by sampling the queue size only at FT arrival instants. Thus, a service differentiation for foreground traffic is possible which does not differ from conventional RED behaviour in the long term average (i.e., in the $ulp$).

Intra-flow loss recovery and control

Temporal sensitivity

Considering a flow with temporal loss sensitivity, paragraph 3.1.1 has shown that a simple, periodic loss pattern enhances the performance of the end-to-end loss recovery. The pattern is not tied to particular packets, therefore a per-flow characterization with the introduced metrics is applicable. In this paragraph we assume that a flow expressed its temporal sensitivity by marking its flow with an alternating pattern of ``+1''/``-1''.

Figures 11 and 12 show the rates for the actual and the estimated burst loss lengths for a three-state ($m=2$) and a two=state ($m=1$, Gilbert) model respectively. We can observe that DiffRED shapes the burst probability curve in the desired way. Most of the probability mass is concentrated at isolated losses (the ideal behaviour would be the occurence of only isolated losses ($k=1$) which can be expressed with $clp=0$ in terms of Gilbert model parameters). With Drop Tail an approximately geometrically decreasing burst loss probability with increasing burst length (Eq. 4) is obtainable, where the $clp$ parameter is relatively large though. Thus, considering voice with temporal loss sensitivity as the foreground traffic of interest, with DiffRED a large number of short annoying bursts can be traded against a larger number of isolated losses and few long loss bursts (which occur when the queue is under temporary overload, i.e., $avg > max_{th}$, Fig. 10).

Figure 11: Comparison of actual and estimated burst loss length rates as a function of burst length $k$: three state run-length-based model

Figure 12: Comparison of actual and estimated burst loss length rates as a function of burst length $k$: two state run-length-based model (Gilbert)

We can see that the three-state model estimation (Eq. 2) reflects the two areas of the DiffRED operation (the sharp drop of the burst loss length rate for $k=2$ and the decrease along a geometrically decreasing asymptote for $k>2$ ). This effect cannot be captured by the two state model (Eq. 4) which thus overestimates the burst loss length rate for $k=2$ and then hugely underestimates it for $k>2$. Interestingly, for Drop Tail, while both models capture the shape of the actual curve, the lower order model is more accurate in the estimation. This can be explained as follows: if the burst loss length probabilities are in fact close to a geometrical distribution, the estimate is more robust if all data is included (note that the run-length based approximation of the conditional loss probability $P(X=m \vert X=m)$ only includes loss run-length occurences larger or equal to $m$: Eq. 1).

Sensitivity due to ADU heterogeneity

In paragraph 3.1.2 we have seen that sensitivity to ADU heterogeneity results in a certain non-periodic loss pattern. Thus, a mechanism at (or near) the sender is necessary which derives that pattern from the voice data. Furthermore, an explicit cooperation between end-to-end and hop-by-hop mechanisms is necessary (Fig. 4).

We use the result of paragraph  3.2 to develop a new packet marking scheme called Speech Property-Based Selective Differential Packet Marking (SPB-DIFFMARK). The DIFFMARK scheme concentrates the higher priority packets on the frames essential to the speech signal and relies on the decoder's concealment for other frames.

Figure 13: SPB-DIFFMARK Pseudo Code

Figure 13 shows the simple algorithm written in a pseudo-code that is used to detect an unvoiced/voiced ($uv$) transition and protect the voiced frames at the beginning of a voiced signal. The procedure analysis() is used to classify a block of $k$ frames as voiced, unvoiced, or $uv$ transition. send() is used to send a block of $k$ frames as a single packet with the appropriate priority (either "+1", "0" or "-1"). As the core algorithm gives only a binary marking decision (protect the packet or not), we employ a simple algorithm to send the necessary ``-1'' packets for compensation (Eq. 5): after a burst of "+1" packets has been sent, a corresponding number of "-1" packets is sent immediately. State about the necessary number of to-be-sent "-1" packets is kept in the event that the SPB algorithm triggers the next "+1" burst before all "-1" packets necessary for compensation are sent. Thus, seen over time intervals which are long compared to the $+1$/$-1$ burst times, the mean loss for the flow will be equal to the "best effort" case (Eq.  5). $N$ is a pre-defined value and defines how many frames at the beginning of a voiced signal are to be protected. Our simulations (Fig. 9) have shown that the range from $10$ to $20$ are appropriate values for $N$ (depending on the network loss condition). In the simulations presented below, we choose $k=2$, a typical value for interactive speech transmissions over the Internet ($20 ms$ of audio data per packet). A larger number of $k$ would help to reduce the relative overhead of the protocol header but also increases the packetization delay and makes sender classification and receiver concealment in case of packet loss (due to a larger loss gap) more difficult.

End-to-end simulation description

Due to the non-periodic loss pattern, we need to explicitly associate a drop probability with a single packet within an end-to-end model. Therefore we use a separate one-state Markov model (Bernouilli model) to describe the network behaviour as seen by each class of packets. "Best effort" packets (designated by "0" in Fig. 14) are dropped with the probability $p_0$, whereas packets marked with ``+1'' and ``-1'' are dropped with probabilities of $p_{+1}$ and $p_{-1}$ respectively. This is a reasonable assumption with regard to the interdependence of the different classes in fact, as it has been shown that DiffRED (Figs. 11 and 12) achieves a fair amount of decorrelation of +1 and -1 packet losses. Nevertheless to include some correlation between the classes we have set $p_{+1} = 10^{-3} \; p_0$ for the subsequent simulations. This should also allow a reasonable evaluation of how losses in the +1 class affect the performance of the SPB-algorithms.

Figure 14: Marking schemes and corresponding network models.

For a direct comparison with SPB-DIFFMARK, we evaluate a scheme where packets are alternatingly marked as being either "-1" or "+1" (ALT-DIFFMARK, Figure  14). We furthermore include related inter-flow loss protection schemes. The first scheme uses full protection (FULL MARK, all packets are marked as "+1"). The SPB-MARK scheme operates similarly to SPB-DIFFMARK, but no ``-1'' packets are sent for compensation (those packets are also marked as ``0''). For comparison we again use a scheme where packets are alternatingly marked as being either "0" or "+1" (ALT-MARK). Finally, packets of pure "best effort" flows are dropped with the probability $p_0$ (NO MARK case in Fig. 14). For the SPB marking schemes the percentage of "+1"- and ``-1''-marked packets respectively is $40.4\%$ for the speech material used. We obtained similar marking percentages for other speech samples. The ALT marking schemes mark exactly $50\%$ of their packets as being ``+1''.

Results

Figure 15 shows the perceptual distortion for the marking schemes dependent on the drop probability $p_0$. The unprotected case ("NO MARK") has the highest perceptual distortion and thus the worst speech quality. The differential marking scheme (SPB-DIFFMARK) offers a significantly better speech quality even when only using a network service which amounts to "best effort" in the long term. Note that the ALT-DIFFMARK marking strategy does not differ from the "best effort" case (which confirms the result of paragraph G.729). SPB-DIFFMARK is also even better than the inter-flow QoS ALT-MARK scheme, especially for higher values of $p_0$. These results validate the strategy of our SPB marking schemes that do not equally mark all packets with a higher priority but rather protect a subset of frames that are essential to the speech quality.

Figure 15: Perceptual Distortion (EMBSD) for the marking schemes and SPB-FEC

The SPB-FEC scheme ([12]), which uses redundancy piggybacked on the main payload packets (RFC 2198, [7]) to protect a subset of the packets, enables a very good output speech quality for low loss rates. However, it should be noted that the amount of data sent over the network is increased by $40.4\%$. Note that the simulation presumes that this additionally consumed bandwidth itself does not contribute significantly to congestion. This assumption is only valid if a small fraction of traffic is voice ([8]). The SPB-FEC curve is convex with increasing $p_0$, as due to the increasing loss correlation an increasing number of consecutive packets carrying redundancy are lost leading to unrecoverable losses. The curve for SPB-DIFFMARK is concave, however, yielding better performance for $p_0 \stackrel{>}{\approx} 0.22$.

The inter-flow QoS ALT-MARK scheme ($50\%$ of the packets are marked) enhances the perceptual quality. However, the auditory distance and the perceptual distortion of the SPB-MARK scheme (with $40.4\%$ of all packets marked) is significantly lower and very close to the quality of the decoded signal when all packets are marked (FULL MARK). This also shows that by protecting the entire flow only a minor improvement in the perceptual quality is obtained. The results for the FULL MARK scheme also show that, while the loss of some of the +1 packets has some measurable impact, the impact on perceptual quality can still be considered to be very low.

Conclusions

In this paper we have characterized the behaviour of a sample-based codec (PCM) and a frame-based codec (G.729) in the presence of packet loss. We have then developed ``intra-flow'' loss recovery and control mechanisms to increase the perceptual quality. While we have tested other codecs only informally, we think that our results reflect the fundamental difference between codecs which either encode the speech waveform directly or which are based on linear prediction. For PCM without loss concealment we have found that it neither exhibits significant temporal sensitivity nor sensitivity to payload heterogeneity. With loss concealment, however, the speech quality is increased but the amount of increase exhibits strong temporal sensitivity. Frame-based codecs amplify on one hand the impact of loss by error propagation, though on the other hand such coding schemes help to perform loss concealment by extrapolation of decoder state. Contrary to sample-based codecs we have shown that the concealment performance of the G.729 decoder may ``break'' at transitions within the speech signal however, thus showing strong sensitivity to payload heterogeneity.

We have briefly introduced a queue management algorithm which is able to control loss patterns without changing the amount of loss and characterized its performance for the loss control of a flow exhibiting temporal sensitivity. Then we developed a new packet marking scheme called Speech Property-Based Selective Differential Packet Marking for an efficient protection of frame-based codecs. The SPB-DIFFMARK scheme concentrates the higher priority packets on the frames essential to the speech signal and relies on the decoder's concealment for other frames. We have also evaluated the mapping of an end-to-end algorithm to inter-flow protection. We have found that the selective marking scheme performs almost as good as the protection of the entire flow at a significantly lower number of necessary high-priority packets.

Thus, combined intra-flow end-to-end / hop-by-hop schemes seem to be well-suited for heavily-loaded networks with a relatively large fraction of voice traffic. This is the case because they neither need the addition of redundancy nor feedback (which would incur additional data and delay overhead) and thus yield stable voice quality also for higher loss rates due to absence of FEC and feedback loss. Such schemes can better accomodate codecs with fixed output bitrates, which are difficult to integrate into FEC schemes requiring adaptivity of both the codec and the redundancy generator. Also, it is useful for adaptive codecs running at the lowest possible bitrate. Avoiding redundancy and feedback is also interesting in multicast conferencing scenarios where the end-to-end loss characteristics of the different paths leading to members of the session are largely different. Our work has clearly focused on linking simple end-to-end models which can be easily parametrized with the known characteristic of hop-by-hop loss control to user-level metrics. An analysis of a large scale deployment of non-adaptive or adaptive FEC as compared to a deployment of our combined scheme requires clearly further study.

Acknowledgments

We would like to thank Wonho Yang and Robert Yantorno, Temple University, for providing the EMBSD software for the objective speech quality measurements. Michael Zander, GMD Fokus, helped with the simulations of the queue management schemes.

Bibliography

1

J. Andren, M. Hilding, and D. Veitch.
Understanding end-to-end internet traffic dynamics.
In Proceedings IEEE GLOBECOM, Sydney, Australia, November 1998.

2

J.-C. Bolot, S. Fosse-Parisis, and D. Towsley.
Adaptive FEC-based error control for interactive audio in the Internet.
In Proceedings IEEE INFOCOM, New York, NY, March 1999.

3

D. Clark and W. Fang.
Explicit allocation of best effort packet delivery service.
Technical Report, MIT LCS, 1997.
http://diffserv.lcs.mit.edu/Papers/exp-alloc-ddc-wf.pdf.

4

R. Cox and P. Kroon.
Low bit-rate speech coders for multimedia communication.
IEEE Communications Magazine, pages 34-41, December 1996.

5

J. Ferrandiz and A. Lazar.
Consecutive packet loss in real-time packet traffic.
In Proceedings of the Fourth International Conference on Data Communications Systems, IFIP TC6, pages 306-324, Barcelona, June 1990.

6

W. Jiang and H. Schulzrinne.
QoS measurement of Internet real-time multimedia services.
In Proceedings NOSSDAV, Chapel Hill, NC, June 2000.

7

C. Perkins, I. Kouvelas, O. Hodson, M. Handley, and J. Bolot.
RTP payload for redundant audio data.
RFC 2198, IETF, September 1997.
ftp://ftp.ietf.org/rfc/rfc2198.txt.

8

M. Podolsky, C. Romer, and S. McCanne.
Simulation of FEC-based error control for packet audio on the Internet.
In Proceedings IEEE INFOCOM, pages 48-52, San Francisco, CA, March 1998.

9

J. Rosenberg.
G. 729 error recovery for Internet Telephony.
Project report, Columbia University, 1997.

10

J. Rosenberg, L. Qiu, and H. Schulzrinne.
Integrating packet FEC into adaptive voice playout buffer algorithms on the Internet.
In Proceedings IEEE INFOCOM, Tel Aviv, Israel, March 2000.

11

H. Sanneck.
Concealment of lost speech packets using adaptive packetization.
In Proceedings IEEE Multimedia Systems, pages 140-149, Austin, TX, June 1998.
ftp://ftp.fokus.gmd.de/pub/glone/papers/Sann9806_Adaptive.pdf.

12

H. Sanneck.
Packet Loss Recovery and Control for Voice Transmission over the Internet.
PhD thesis, GMD Fokus / Telecommunication Networks Group, Technical University of Berlin, October 2000.
http://sanneck.net/research/publications/thesis/Sann0010_Loss.pdf.

13

H. Sanneck and G. Carle.
A queue management algorithm for intra-flow service differentiation in the "best effort" Internet.
In Proceedings of the Eighth Conference on Computer Communications and Networks (ICCCN), pages 419-426, Natick, MA, October 1999.
ftp://ftp.fokus.gmd.de/pub/glone/papers/Sann9910_Intra-Flow.pdf.

14

H. Sanneck, N. Le, and A. Wolisz.
Efficient QoS support for Voice-over-IP applications using selective packet marking.
In Special Session on Error Control Techniques for Real-time Delivery of Multimedia data, First International Workshop on Intelligent Multimedia Computing (IMMCN), pages 553-556, Atlantic City, NJ, February 2000.
ftp://ftp.fokus.gmd.de/pub/glone/papers/Sann0002_VoIP-marking.pdf.

15

H. Schulzrinne, J. Kurose, and D. Towsley.
Loss correlation for queues with bursty input streams.
In Proceedings ICC, pages 219-224, Chicago, IL, 1992.

16

D. Sisalem and A. Wolisz.
LDA+ TCP-friendly adaptation: A measurement and comparison study.
In Proceedings NOSSDAV, Chapel Hill, NC, June 2000.

17

International Telecommunication Union.
Coding of speech at 8 kbit/s using conjugate-structure algebraic-code-excited linear-prediction (CS-ACELP).
Recommendation G.729, ITU-T, March 1996.

18

International Telecommunication Union.
Objective quality measurement of telephone-band (300-3400 Hz) speech codecs.
Recommendation P.861, ITU-T, February 1998.

19

M. Yajnik, J. Kurose, and D. Towsley.
Packet loss correlation in the MBone multicast network: Experimental measurements and markov chain models.
Technical Report 95-115, Department of Computer Science, University of Massachusetts, Amherst, 1995.

20

M. Yajnik, S. Moon, J. Kurose, and D. Towsley.
Measurement and modelling of the temporal dependence in packet loss.
Technical Report 98-78, Department of Computer Science, University of Massachusetts, Amherst, 1998.

21

W. Yang and R. Yantorno.
Improvement of MBSD by scaling noise masking threshold and correlation analysis with MOS difference instead of MOS.
In Proceedings ICASSP, pages 673-676, Phoenix, AZ, March 1999.

Footnotes

... ADUs

Application Data Unit: the unit of data emitted by a source coder such as a video or voice frame.

... behavior

While we have obtained results which confirm the shape of the ``overall utility'' curve shown in Fig. 1, clearly the utility functions of the +1/-1 ``sub''-flows and their relationship are more complex and only approximately additive.

... (LSP

LSPs are another representation of the linear prediction coefficients.

... G.729

Two G.729 frames are contained in a packet.

... degradation

While on one hand we see that $SNR$ measures often do not correlate well with subjective speech quality, on the other hand the large differences in the $SNR$ -threshold-based resynchronization time clearly point to a significant impact on subjective speech quality.

... respectively

We only discuss the results qualitatively here to give an example how an intra-flow loss control algorithm performs and to show how loss models can capture this performance. Details on the simulation scenario and parameters can be found in [12].

... assumption

The appropriateness of the simple end-to-end modeling used has been investigated in [12] with discrete event simulation using a multi-hop topology and detailed modeling of foreground and background traffic sources.

... quality

We have also performed informal listening tests which confirmed the results using the objective metrics.

... redundancy

We also used the G.729 encoder for the redundant source coding.

Copyright © 2001 ACM.