Congestion Control in Unicast Applications

Initial buffering in receivers smoothes away small fluctuations in throughput and transmission delay. When the data rate throughput change is such long in duration or big in bitrate that re-ceiver buffers are in danger of under- or overflowing, senders should adjust the transmission bitrate to the available throughput. Sources of throughput changes and detection methods for them are reviewed in Section 4.2.1.

Abrupt and substantial throughput drops, such as those caused by cell handovers in mobile reception, pose a challenge to continuous media rendering. Even if senders react to detected throughput drops immediately by adjusting the transmission bitrate, client buffer may drain in the meanwhile and cause discontinuous playback. In order to counter throughput drops proactively, robust packet scheduling methods can be used, as discussed in Sec-tion 4.2.2.

When congestion in fixed packet-switched network is detected, senders are often ex-pected to behave similarly to the congestion control algorithm of the Transmission Control Protocol (TCP), which adjusts the sending rate by an Additive Increase Multiplicative De-crease (AIMD) rate control algorithm [18]. Additive inDe-crease refers to the fact that the trans-mission rate is raised incrementally if no packet losses are detected, and multiplicative de-crease refers to halving the transmission rate once every round-trip time as a response to a packet loss. It is obvious that when a sender encodes video in real-time and there is only one receiver, the bitrate control algorithm of the encoder can be used for data rate adjustment.

Otherwise, methods manipulating coded bitstreams, such as stream thinning and switching, should be used. A brief review of stream thinning and switching methods is given in Sec-tion 4.2.3.

Error Resilience in H.264/AVC Video Communication 33 4.2.1. Sources and Detection of Throughput Changes

There are three main sources for throughput drops in streaming over heterogeneous networks:

congestion in a best-effort network, decrease of a service class in networks providing guaran-teed quality of service (QoS), and cell handovers in mobile networks. Each type of a through-put drop requires specific handling, even though the underlying processing techniques remain the same for all.

RTCP receiver reports provide some means for detecting throughput changes. When congestion is the only source of packet losses, the packet loss counters included in RTCP re-ceiver reports give an indication of congestion. RTCP rere-ceiver reports also include the interar-rival jitter field, which represents the statistical variance of packet interarinterar-rival times relative to RTP timestamps. Increased interarrival jitter can be of a sign of congestion and enables the sender to react before packet losses happen. [117]

Many of the currently used packet-switched mobile networks, such as GPRS, offer a best-effort service. The use of RTCP receiver reports for throughput change detection is unre-liable in best-effort mobile networks, because the wireless link is an additional source of packet losses. Thus, senders cannot conclude based on RTCP receiver reports whether an in-creased packet loss rate is due to congestion or more severe radio conditions. Moreover, the use of the acknowledged mode in the link layer may cause substantial end-to-end delay varia-tion, which makes methods based on interarrival jitter inoperable. [36]

In order to provide means to differentiate between congestion and harsh radio condi-tions, an RTCP extended report with receiver buffer status indicacondi-tions, also known as RTCP APP packet with client buffer feedback (NADU APP packet), has been specified in the 3GPP packet-switched streaming service [2]. The signaling enables the sender to reconstruct the ex-act status of the receiver buffer and derive the delay prevailing in the network. Reviews of the bitrate adaptation feature of 3GPP packet-switched streaming are provided in [26] and [27].

An example of signaling for a throughput change when guaranteed QoS is in use can also be found in the 3GPP packet-switched streaming service [2], which specifies the 3GPP-Link-Char header to be used with certain RTSP methods for signaling of the guaranteed radio link bandwidth.

Cell handovers may cause data rate throughput drops that are typical for mobile net-works. When it comes to the detection of a handover in the application layer of real-time mul-timedia applications, Kampman and Baldo proposed to use the NADU APP packet for hand-over detection [77]. While the proposed technique detects handhand-overs accurately, the detection can only operate after the handover is over, as no RTCP extended report is conveyed during the handover. Therefore, Bouazizi et al. proposed handover detection based on the expected reception interval of RTCP receiver reports [14].

4.2.2. Robust Packet Scheduling

Robust packet scheduling techniques are applicable when recipients buffer data initially be-fore the start of media playback. The techniques can be applied for minimizing the impact of

abrupt network throughput changes, such as cell handovers, and for selecting an optimal trun-cation path for stream thinning. The idea of robust packet scheduling is as follows: Coded slices and coded slice data partitions that are subjectively the most important are sent earlier than their decoding order indicates, whereas coded slices and coded slice data partitions that are subjectively the least important are sent later than their natural decoding order indicates.

Consequently, any temporary decrease in the network throughput might cause the least impor-tant data to arrive too late for decoding while sufficient amount of the most imporimpor-tant data would be readily buffered in the receiver to compensate for the throughput drop. Moreover, if a piece of the most important data is lost during transmission, it is more likely that the piece could be retransmitted and received before its scheduled decoding or playback time compared to the least important data. [54]

The algorithms for robust packet scheduling can be categorized into two classes: heu-ristic and rate-distortion-optimized algorithms. Two examples of the heuheu-ristic algorithms are given in this paragraph, whereas the rate-distortion-optimized algorithms are introduced in the next paragraph. Miao and Ortega proposed an algorithm called Expected Runtime Distortion Based Scheduling (ERDBS) [93][94], which derives the importance and scheduling of each packet in run-time based on the scalability layer which the packet contains data for, the fact whether any data dependent on the packet has already been transmitted and received cor-rectly, and the probability of the packet to reach the destination before the decoding and play-back time of the packet. Kang and Zakhor [78] proposed an algorithm that assigns an impor-tance factor to packets and pictures carried in the packets as an increasing function of the pic-ture decoding order relative to the previous intra picpic-ture starting a GOP. Packets are sent in increasing order of the importance factor within a certain time window relative to the current playback time in the recipient. The algorithm is integrated with feedback from the recipient indicating which packets have been correctly received. Moreover, the algorithm can be ap-plied together with data partitioning by weighting motion and texture packets differently. In their later work [79], Kang and Zakhor developed a rate-distortion-optimized version of the heuristic algorithm [78]. The algorithm uses statistical video traffic models and considers channel bitrate fluctuations.

Chou and Miao developed rate-distortion-optimized algorithms for robust packet scheduling in streaming applications in their technical report [19], later published as a journal paper [21]. The algorithm optimizes the use of time and bandwidth resources by minimizing a Lagrangian rate-distortion cost function. The paper concludes that rate-distortion-optimized streaming of an entire presentation can be solved by focusing on the error-cost optimized transmission of a single data unit in isolation. The presented algorithm can be applied to vari-ous streaming scenarios, including sender-driven and receiver-driven, with and without feed-back, with and without retransmission and forward error correction, as well as best-effort and guaranteed quality of service. The work of Chou and Miao has been used as the basis for sev-eral enhancements and extensions to new applications, such as considerations for multiple playback deadlines and accelerated retroactive decoding [76], multi-path transmission [17], and rate-congestion optimization considering congestion avoidance on bottleneck links [121].

Error Resilience in H.264/AVC Video Communication 35 In order to facilitate robust packet scheduling, the transmission system has to provide

means for sending data out of its decoding order and recovering the data decoding order in receivers. The RTP sequence number is required to be incremented by one for each sent RTP packet, therefore providing the capability of recovering the transmission order of packets in receivers. However, RTP includes no mechanism for recovery of a correct decoding order if data is not transmitted in decoding order. A mechanism for the RTP payload format of H.264/AVC allowing any data transmission order and recovery of the correct decoding order in receivers is presented in Section 6.2.2. The presented mechanism includes signaling ena-bling receivers to allocate a data reordering buffer having a sufficient size and applying a cor-rect amount of initial data buffering. Section 6.3.1 discusses how robust packet scheduling can be applied for H.264/AVC.

4.2.3. Stream Thinning and Switching

Stream thinning refers to omission of certain coded data units, such as non-reference pictures and the least important scalability layers, from the transmitted stream. Even non-scalable bitstreams can be thinned as explained next. A known method in current streaming systems to cope with drastically dropped channel throughput is to transmit intra-coded pic-tures only. When the network throughput is restored, inter-coded picpic-tures can be transmitted again from the beginning of the next GOP. Generally, any chain of inter-coded pictures can be safely disposed, if no other picture is predicted from them. Consequently, inter-coded pictures at the tail of a GOP can be removed without affecting the decoding of any previous or subse-quent picture. In general, priority partitioning methods, reviewed in Section 4.1, can be used to select the order according to which parts of the bitstream are omitted from the transmitted stream when the channel throughput is not sufficient. The sub-sequence technique, presented in Chapter 6, can be used to increase the number of priority classes and hence provide finer bitrate adaptation steps compared to what can be achieved with conventional non-hierarchical temporal scalability.

If stream thinning does not provide a big enough dynamic range for bitrate adjustment, the server should switch to a different version of the same content coded for a bitrate that is close to the network throughput. Switching to a different bitstream can naturally be done at any random access point. In order to respond to a need for adjusting bitrate faster and avoid the compression penalty of frequent intra pictures, there have been studies how stream switch-ing could be done startswitch-ing from non-intra pictures. Färber and Girod proposed S frames that are inter-coded frames used only when switching from a first stream to a second stream [39].

S frames are encoded with a small quantization step and make the decoded S frame close but typically not identical to the corresponding decoded picture of the second stream. H.264/AVC includes the feature known as SI/SP pictures [80], which can be used similarly to S frames but provide identical decoded picture after switching compared to decoding of the stream from the beginning. Identical decoded pictures are obtained with the cost of additional transform and quantization steps in the decoding process for SI/SP pictures both in the primary streams

and SI/SP pictures used for switching only. However, the SI/SP feature is not included in the Baseline or High profile and therefore not commonly used.