«End-to-End QoS Provision over Heterogeneous IP and non IP Broadband Wired and Wireless Network Environments A dissertation submitted in satisfaction ...»
Unless ﬁrewalls force them to, streaming media systems do not rely on TCP but implement their own, application-layer transmport mechanisms. This allows for protocols that are both network adaptive and media aware. A transport protocol may determine, for example, when to retransmit packets for error control and when to drop packets to avoid network congestion. If the protocol takes into consideration the relative importance of packets and their mutual depedencies, audio or video quality can be greatly improved.
The media server can implement intelligent transport by sending the right packets at the right time, but the computational resources available for each media stream are often limited because a large number of streams must be served simultaneously. Much of burden of an eﬃcient and robust system is therefore in the encoder application, which, however, cannot adapt to the varying channel conditions and must rely on the media server for this task. Rate scalabel representations are therefore desirable to facilitate adaptation to varying network throughput with-out requiring computation at the media server. Switching among bit streams encoded at diﬀerent rates is an easy way to achieve this task, and this method is widely used in commercial systems. Embedded scalable representation, as discussed in previous chapter for video, are more elegant and are preferable, if the rate-distortion penalty often associated with scalable coding can be ketp small.
2.4.1 Rate-Distortion Optimized Streaming
Let us assume that a media server has stored a compressed video stream that has been packetized into data units. Each data unit l has a suize in bytes Bl and a deadline by which it must arrive at the client in order to be useful for decoding.
The importance of each data unit is captured by its distortion reduction δDl, a value representing the decrease in distortion that results if the data unit is decoded. Often, distortion is expressed as mean-squared error, but other distortion measures might be used as well.
Whether a data unit can be decoded often depends on which other data units are available. In the RaDio framework, these inter-dependencies are expressed in a directed acyclic graph. An example dependency graph is shown for SNR-scalable video encoding with Intra (I), Predicted (P), and Bidirectionally predicted (B) frames as shown in Figure 2.4.1. Each square represents a data unit and the arrows indicate the order in which data units can be decoded.
The RaDio framework can be used to choose an optimal set of data units Figure 2.4: A directed acyclic graph captures the decoding dependecies for an SNR-scalable encoding of video with I-frames, P-frames, and B-frames. Squares represent data units and arrows indicate decoding order.
to transmit at successive transmission opportunities. These transmission opportunities are assumed to occur at regular time intervals. Because of decoding dependencies among data units, the importance of transmitting a packet at a given transmission opportunity often depends on which packets will be transmission decisions based on an entire optimized plan that includes anticipated later transmissions. Of course, ato keep the system practical, onlu a ﬁnite time horizon can be considered.
The plan governing packet transmissions that will occur within a time horizon is called a tramission policy, π. Assuming a time horizon of N transmission opportunities, π is a set of lenght-N binary vectors πl, with ine such vector for each data unit l unider consideration for transmission. In this representation, the N binary elemets of π indicate wheter, under the policy, the data unit l will be transmitted at each of the next N transmission opportunities. The policy is understood to be contigent upon future acknowledgments that might arrive from the client to indicate that the packet has been received. No further transmissions of an acknoledgment data unit l are attempted, even if π speciﬁes a tranmission for a future time slot.
Each tramission policy leads to its own error probability, (πl ), deﬁned as the probability that data unit l arrives at the client late, or not at all. Each policy is aslo associated with an expected number of times that the packet is transmitted unider the policy, ρ(πl ). The goal of the packet scheduler is to ﬁnd a transmission policy π with the best trade-oﬀ between expected transmission rate and expected reconstruction distortion. At any transmission opportunity
the optimal π minimizes the Langragian cost function:
In the aﬀorementioned formulation, delays and losses experienced by packets transmitted over the network are assumed to be statistically independent. Packet loss is typically modeled as Bernoulli with some probability, adn the delay of arriving packets is often assumed to be a shifted-Γ distribution. Expressions for (πl ) and ρ(πi ) can be derived in terms of the Bernouli loss probabilities, the cumulative distribution functions for the Γ-distriguted delayes, the transmission poliocies and transmission histories, and the data units’ arrival deadlines. These derivation are straightforward, but because the resulting expression are cumbersome, thre are ommitted here.
The scheduler re-optimizes the entire police π at each transmission opportunity to take into account new information since the previous transmission opportunity and then exectues the optimal π for the current time. An exhaustive searc to ﬁnd the optimal π is general nto tractable; the search space grows exponentially with the number of considered data units, M, and the lenth of the policy vector, N . Even though rates and distortion reductions are assumed to be additive, the graph of packet dependencies leads to interactions, and an axhaustive search would have to consider all 2M N possible policies.  overcome this problem by using conjugate direction search. Their Iterative Sensitivity Adjustment (ISA) alogrithm minimizes 2.1 with respect to the policy πl of one data unit while the transmission policies of other data units are held ﬁxed. Data units’ policies are optimized in round-robin order until the Langragian cost converges to a minimum.
Rewriten in termms of the transmission policy of one data unit, equations 2.1, 2.2 and 2.3 become:
the data unit size Bl and Sl, a term that expresses the sensitivity of the overall expected distortion to the error probability of data unit and incorporates the error probabilities of the onter dta units tah l depends on. The sensitivity Sl changes ith the iteration of the proposes algorithm to take into account the optimized policy for the other data units.
2.4.2 Congestion-Distortion Optimized Scheduling
Radio streaming and it various extensions descrﬁbed do not consider the eﬀect that transmitted media packets mya have on the delay of subsequently transmitted packets. Dealy is modeled as a random variable with a parameterized distribution; parameters are adapted slowllu according to feedback information.
IN the case when the media stream is transmitted at a rate that is neglible compared to the minimum link speed on the path from server to client, this may be an acceptable model. In the case where there is a bottlenck link on the path from server to client, however, packet delays can be strongly aﬀcted by self-congestion resulting from previous transmissions.
Authors in  propose a congestion-distortion optimized algorithm, which takes into account the eﬀect of transmitted packets on delays. The scheme is intended to achive an R-D performance similar to RaDio streaming but specifically schedules packet transmissions in a way that uields an optimal trade-oﬀ between reconstruction distortion and congestion, measured as average delay on the bottlenecked link. As with RaDiO, tranmissin actions are chosen at discrete transmission opportunities by ﬁnding an optimal policy over a time horizon. However, in this proposed framework, the optimal policy minimizes the Langragian cost D + λ, where D is the expectged distortion due to the policy and ∆ is the
expected end-to-end delayt, which measures the congestion.
The proposed framework’s channel model assumes a succession of high-badnwidth links shared by many users, followd by a bottleneck last hop used only by the media stream under consideration. CoDiO needs to know the capacity of the bottleneck, which can be estimated, for example, by transmitteing back-to-back packets . The overall delay experienced by packets is captured by a gammapdf that is dynamically shifted by an extra delay that models the self-inﬁcterd backlog at the bottleneck. Since the bottleneck is not shared and its capacity is known, the backlog, can be accurately estimated. This channel model is used to calculae the expected distortion D due to packet loss and the xpected end-to-end delay ∆.
2.5 Inter-domain techniques for providing traﬃc diﬀerentiation at the network level This section reviews basic technologies that can oﬀer QOS support in both wired and wireless network domains. Particularly, the relevant technlogies for IP, DVB, 3G and 802.11 networks are outlined.
2.5.1 IP Domain - Diﬀerentiated Services The Diﬀerentiated Services  (DiﬀServ) framework aims to provide service differentiation within backbone IP networks. It provides QoS only to aggregated traﬃc classes rather than to speciﬁc ﬂows, like IntServ, without the use of signalling mechanism.
Essentially, on entry to a network, packets are placed into a broad service group by a classiﬁcation mechanism that reads the DiﬀServ Code Point  (DSCP) in the IP packet header and the source and destination address.
A number of diﬀerent classes has been deﬁned. These are the Expedited Forwarding [?] (EF) class, which aims to provide a low-jitter and low-delay service.
Users must operate at a known peak rate and packets will be discarded if users exceed their peak rate. The Assured Forwarding  (AF) classes are intended for delay tolerant applications. Here, the guarantees simply imply that the higher QoS classes will perfom better, faster delivery and lower loss propability, than the lower classes. Furthermore, network operators are also at liberty to deﬁne their own per-hop behaviors, note that the use of these behaviors requires packet remarking on network boundaries.
One DiﬀServ class may be used by many ﬂows. Any packet within the same class must share resources with all other packets in that class. Packets are treated on a per-hop basis by traﬃc conditioners. The main issue with regard to QoS service provision is the handling of packets from aggregated ﬂows through ﬁve basic network components of the DiﬀServ architecture, which are, the Classiﬁer that seperates submitted traﬃc into diﬀerent classes, the Traﬃc Conditioner that forces the traﬃc to conform to a proﬁle, the Queue Management that controls the status of queues under congestion conditions, the Scheduler that detetermines when the packet will be forwarded and ﬁnally the Admission Control that is usually used in absolute service diﬀerentiation [14, 15].
DiﬀServ removes the InteServ’s per-ﬂow state and scheduling that leads to scalability problems. However, it provides only a static QoS conﬁguration,typically through Service Level Agreement, as there is no signaling for the negotiation of QoS.
2.5.2 DVB Domain - Bandiwdth Management In order to transmit IP trafc over a DVB network, the IP packets need to be encapsulated in MPEG-2 TS packets. The encapsulation of IP data into MPEGTS packets follows the Multi Protocol Encapsulation (MPE) standard , .
According to the standard, depending on the performed encapsulation mode, the encapsulation process adds an overhead that ranges from 16 bytes to 44 bytes.
The resulting TS, being the outcome of the encapsulation, is subsequently multiplexed with other MPEG-2 TSs, which might either contain IP data or MPEG-2 Digital TV services. The outcome is a multiplexed TS containing various services. Each MPEG-2 TS is assigned a Program Identication (PID) value in order to be discriminated among other MPEG-2 TSs. After its multiplexion, TS is modulated, up-converted and transmitted. In the reception, the received DVB signal is down-converted, demodulated and then ltered (using the PID value) in order each receiver to take its own data.
The above multiplexing method does not oﬀer any kind of trafc diﬀerentiation. Dealing with trafc diﬀerentiation issue, a DVB network can apply on queues containing 188 byte long MPEG-2 TS packets a BM technique. This technique is based on the dynamic uplink bandwidth reallocation into a number of independent virtual channels according to a predened set of priority policies. Figure 2.5.2 depicts the bandwidth slicing principle of a DVB uplink into a number of virtual IP channels, each one supporting a specic bit rate that can be assigned to a diﬀerent service. The assignment of an IP ow at a virtual channel is achieved through a ltering mechanism, who is able to monitor trafc and based on some pre-dened lters (IP source and destination addresses, source and destination ports, DSCP, TOS, protocol type etc) to encapsulate that trafc to a specic virtual channel.
The actual implementation of a BM technique requires two modules, i.e, an
encapsulator and a statistical multiplexer. Generally speaking the encapsulator is responsible for monitoring the IP data trafc and based on the discussed lters (packet identiers) to choose the packets that will be delivered and the statistical multiplexer undertakes to smooth out peaks of the individual MPEG-2 TSs within the aggregated output transport.
2.5.3 UMTS Domain - UMTS QoS architecture The main goal of UMTS QoS architecture is to provide data delivery with appropriate end-to-end, from user equipment (UE) to UE, quality of service guarantees and is based on a layered bearer1 service architecture .
The end-to-end bearer service is constituted from three basic components:
A bearer service is a type of telecommunication service that provides the capability of transmission between access points