NOTE03L07: INTRODUCTION TO MULTIMEDIA COMMUNICATION

Transcription

1 NOTE03L07: INTRODUCTION TO MULTIMEDIA COMMUNICATION Some Concepts in Networking Circuit Switching This requires an end-to-end connection to be set up before data transmission can begin. Upon setup, the path will be dedicated to the connection until the end of data transmission. Implications: Round trip delays for sending the request signal and propagating the acknowledgement back to the sender Upon setup, the only delay now will be due to the physical data transmission speed of the medium No congestion after setup, though there could be congestion before setup Advance static reservation could lead to waste of resources. Packet Switching No fixed path is established between sender and receiver. Each data to be sent is divided into blocks called packets. Each packet is assigned a sequence number, and then treated as an independent entity. Thus, packets that belong to the same data can travel over different hops in the network before reaching the destination. The packets can also arrive at the destination at different times, and not necessarily in the order that they were sent. Using the packet numbers, the packets are re-assembled at the receiver to produce the original data. Pros No call connection time No dedicated network path Dynamic use of network resources Cons Delays due to congestion, and time spent at each hop Buffer space may be required to keep the packets before they are transmitted Possible packet loss Packet re-assembly time More header information could lead to wastes, since we could have many packets Statistical multiplexing In deterministic multiplexing, each connection is allocated its maximum bandwidth requirement. Under statistical multiplexing, many variable bit-rate connections can share a link with a capacity less than the sum of their maximum bit rate requirements. This is based on the idea of bandwidth-on demand, i.e. when an application is not using some of its allocated bandwidth, some others can use it. When the data streams are bursty and independent, some streams will be at low bit rates, while others will be at high bit rates. Thus, statistical multiplexing will provide a more efficient use of network resources in such environments. Multicast and Broadcast In broadcast, one sender transmits data to all receivers. In multicast, one sender can transmit data to only a group of receivers. The constraint is that, the same data should not pass through a given network link more than once. The objective here is more efficiency and better use of network resources. The problems are reliability (if the single copy has a problem, no one can receive), and a possible bottleneck at the final distribution node. 1

2 Performance Parameters for MM Networks Multimedia Applications The multimedia applications with the most stringent requirement on the network often involve one of the following types: Streaming stored multimedia data Streaming live data Real-time interactive applications Network Speed and Bandwidth User access bandwidth at user access points Network aggregate bandwidth at network exchanges Factors that affect network bandwidth Underlying physical transmission medium Intermediate nodes capacity, failure, speed Congestion Buffer capacity Flow control Error Rates Bit error rates Packet error rates Frame error rates Delay Absolute delays End-to-end delay Round-trip delay Delay variation Efficient sharing for network resources Packet buffers, network bandwidth, etc. should be shared efficiently and fairly. Packet size - the amount of header information could affect network efficiency. Number of retransmissions Need for statistical multiplexing Network scalability Distance scalability LAN, MAN, WAN Bandwidth scalability network bandwidth should be able to grow without requiring changes in network or transport protocols User scalability - one basic problem with most LANs Multicast Ability Send data efficiently to a group of users 2

3 Characteristics of MM Traffic Multimedia traffic usually consists of long streams of data, generated by a multimedia source, such as a video camera. MM traffic streams can be characterized by the following: Variation of throughput with time Constant Bit-Rate Traffic Here, the data traffic is constant over a given time. This usually results from applications that generate outputs at a constant rate, example, applications using uncompressed video. The networks will thus need to support these data streams at a constant bit rate too. Variable Bit-Rate Traffic Here, the data traffic varies with time. The variation usually occurs in the form of short bursts of heavy traffic and random periods of relative inactivity. Compressed video streams typically exhibit this kind of characteristics. VBR video traffic is primarily due to two factors: the compression algorithms used, and the variation in video scene content. So here, we may need to consider such factors as the peak, average, and variation in data rates, and burst intervals. Time dependency Some applications require real time transfer of the data streams. This is usually the case with collaborative/conversational applications, such as video conferencing. Thus the end-to-end delay and the delay jitter must be kept as small as possible. Bi-directional symmetry A network is said to be asymmetric, if the traffic in one direction is much more than the traffic in the other direction. This is typically the case in VoD applications, whereby the data traffic from the server to a subscriber is usually much larger than the traffic from an individual client to the VoD server. In some other applications, the traffic may exhibit bidirectional symmetry, in that about the same amount of traffic is observed to and fro a given participant. Example here is in teleconferencing. Networks Suitable for MM Communication From the above discussions, we can expect the following from a communication network for multimedia data: The network should provide large individual access bandwidth The network should be based on packet switching and statistical multiplexing The network should provide guarantee for throughput, error rates, delay and delay jitter. The network should be scalable in terms of distance, bandwidth, and number of users The network should have capability for multicast. 4 Principles for Providing QoS Support over Unreliable Networks Packet classification Isolation among traffic flows Efficiency sharing of network resources Call admission process 3

4 Networking Technologies Ethernet Ethernet is a widely used LAN, because of its generally low cost. It is based on the CSMA/CD (carrier sense multiple access with collision detection) medium access protocol. It forms the basis for the IEEE LAN standard. The access control for the CSMA/CD is as follows: A station listens to know if the channel is free. When the channel is free, the station transmits. If there is a collision, the station stops transmitting, and sends a jam signal to warn other stations of the collision. The station then waits a random delay, before starting the process again. In case of successive collisions, the station will give up transmission after a certain number of trials, usually 16. Basic Characteristics Data rate 10Mbps Maximum distance between stations 2.5km Maximum number of stations 1024 MAC Protocol CSMA/CD Analysis of the CSMA/CD protocol The random delay is usually an integer multiple of twice the maximum propagation time, τ. Given d = distance between the farthest nodes, and s= signal propagation speed τ = d s For a 1000m cable, τ ~ 5µs The network efficiency is given as the fraction of time that the nodes in the network can transmit new packets under a heavy load imposed by all the stations. Let n > 1 be the number of stations The nodes will attempt to transmit at discrete times, with slot duration = 2τ. Let p= probability that each node transmits during an idle (contention) slot Let p s = probability that a station successfully acquires the channel. This is the probability that exactly one station transmits during the contention slot. Then, p s = n. p n ( 1 p) 1 The value of p that maximizes p s is simply given by dps = n. dp This evaluates to n 1 n 2 ( 1 p) n. ( n 1) p( 1 p) = 0 1 p = n Then, we can find the maximum probability of successful transmission as p n 1 s = 1 1 n 1 e as n The mean number of slots per contention interval is given by: 4

5 n c = i= i= 0 i. p s i ( 1 p ) s 1 1 = p With a slot of size 2τ, we then have a mean contention interval transmission. Network efficiency s actual time taken to transmit η = total time needed n c. 2τ = 2τe, assuming we use the optimal probability of a packet Tp = T + T where T p the actual time used to transmit a packet T c is the time wasted due to contention. Then p c η = T p Tp 1 = + 2 τe 1+ 2τ e T p T p usually depends on the packet size and the available bandwidth, while τ depends on the signal propagation speed and the maximum distance between stations. We can then see that, with the Ethernet, the more bandwidth we provide or the more distance we wish to cover, the worse the network utilization. Therefore, adding more bandwidth to support more users will not help much. Ethernet and MM communication Probabilistic MAC protocol means that a station can wait for an arbitrarily long time before access is granted. The worst case delay is thus is unbounded No priorities means that it is not suitable for real-time traffic Performance depends critically on cable length which determines the propagation delay, and is independent of input data rate. At high loads, collision problem becomes a major headache, and thus many users cannot be supported. Provides multicast ability The Ethernet has a fundamental limitation that makes it difficult to use for large-scale MM communication. Token Rings The token ring is a collection of ring interfaces connected by point-to-point links. It makes use of a token which is a special bit pattern that circulates around the ring whenever all the stations are idle. The token ring forms the basis for the IEEE LAN standard. A station that wants to transmit must first capture the token and remove it from the ring. This is done by changing the token from idle to busy. The station then transmits (usually for a stated token holding time) while the other stations wait. When the station finishes transmitting, it re-injects the token into the ring. Other stations can then re-capture the free token and transmit their own data. This means that there will be no collision, since there is only one token. Basic Characteristics Data rate 16Mbps Maximum distance between stations 2.5km Maximum number of stations 250 MAC Protocol Token Access Method Performance Provides priorities Throughput and efficiency increases with load 5

6 Frame length is arbitrary, and depends on the token holding time. Also large packet sizes could lead to packetization delay. Also, some delay is introduced in waiting for the token Uses round robin and thus fairer than the Ethernet protocol Known upper bound on channel access time Requires ring maintenance for lost tokens, orphan frames, gabbled frames, etc Overall, with bounded delay, availability of priorities and better network utilization, this provides a better option than the Ethernet for MM support. However, the 16Mbps bandwidth is still too small for large user communities, and the delay may be bounded, but still long. Fiber Distributed Data Interface (FDDI) Basic Characteristics Data rate 100Mbps Maximum distance between stations 200km Maximum number of stations 1000 MAC Protocol Token Access Method FDDI is an extension of the token ring concept. It consists of two fiber rings, one transmits in the clockwise direction, while the other transmits in the counter clockwise direction. If one breaks, the other can serve as a backup. If the two rings break at the same time, they can be joined into a single (longer) ring. Thus, the FDDI provides a better reliability than the usual token ring. Like in the token ring, before a station can transmit, it must first capture the token. It then sends its frames and removes them as they cycle back. The FDDI provides both asynchronous and synchronous service. Asynchronous service is used for the ring traffic that is not very delay-sensitive. Here, data is sent only when the token holding rules allow. Basically, this is used when some stations do not need to use the bandwidth allocated to then for the synchronous service. In synchronous service, each station is guaranteed a certain proportion of the FDDI bandwidth. A station can transmit frames any time as long as it does not exceed the synchronous bandwidth allotted to it. FDDI operations are regulated by the use of various timers. The Target Token Rotation Time (TTRT) is a time that is negotiated by all the stations. It represents an estimate of the time it will take each station to receive a token. The Token Rotation Timer (TRT) times the receipt of tokens. It checks the time that has elapsed since the last time the station received a token. It is initialized to the TTRT. A station whose TRT counts down twice (i.e. waits for 2*TTRT) without seeing a token assumes that the token is lost and starts a token initialization process. Thus the maximum delay is 2*TTRT. The Token Holding Time (THT) is another timer that controls the length of time that a station can transmit asynchronous frames. Whenever a station captures a token, its TRT will be copied to the THT. Thus THT TTRT. Usually, for synchronous frames, a station can transmit data for an average of TTRT/n time units, where n is the number of stations. After the expiry of the TTRT/n time units, the station can still transmit asynchronous frames if the existing token rules allow i.e. if THT > 0. Performance Synchronous service can guarantee delay bounds and minimal bandwidth for each connection There is however need for buffers at the sender, since token may take a long time to arrive. Some buffers may also be needed at the receiver to smoothen out delay variations FDDI-II divides the 100Mbps into 16 wide band channels of 6.144Mbps, and provides isochronous service. Yet, reservation inherently comes with some waste of resources. 6

7 Overall, FDDI provides a viable option for MM communication. However, at 100Mbps, less than 100 MPEG-1 video streams can be supported at the same time. Asynchronous Transfer Mode (ATM) The ATM provides a switching and multiplexing technique that is independent of the transmission medium or the transmission technology. It provides standard bandwidths at 155.5Mbps and 622Mbps. It has been made popular by its choice as the backbone network for the ISDN. The basic objectives that motivated the introduction of the ATM were: Integration of different data types, such as voice, non-voice audio, video, images, and data services To keep up with the high-speed of transmission by minimizing the complexity in terms of switching and buffer management at the intermediate nodes. The main idea was then to Carry the connection identifiers along with each data unit (called cells) in every time slot. This implies an elaborate mechanism for the headers of each unit. Keep the size of the time slot small enough, so that in the event of a data (cell) loss, the effect will be minimal Use of fixed-size cells since this will be much easier to handle than variable length cells. In the ATM, cell slots can be used independently by any channel. Thus, one major feature of the ATM is bandwidth on demand whereby each station can use as much bandwidth as it needs, subject to a maximum bandwidth. (The maximum is determined by the access bandwidth) Unused slots can be utilized by other connections thus bandwidth is shared efficiently among those that need it. An ATM network performs operations based only on the cell header. The header contains the necessary information for transporting the data from the sender to the receiver while meeting the QoS specification. The ATM does not know and is not worried by the actual information content of a cell. The part of the cell that contains the data is called the payload. ATM Cell Header Determinants of Cell Size Overhead ratio of the header to the total cell size (i.e. header plus payload). For better network efficiency, we should have small overhead. Cell assembly delay akin to the usual packetization delay. The larger the cell size, the higher the delay, thus we should have smaller cell sizes. Multiplexing delay this is the delay suffered by the cells at the network links due to the multiplexing of the cells. The delay depends on how fast the link can service each request, the number of requests, and the size of the data cells. The larger the size, the more the delay. Forwarding delay the time difference between the instant the switch receives the first bit of the cell to the time it starts forwarding the cell to the outgoing link. For some types of network, they have to receive all the data cells before they can forward it. For these networks, there will still be a forwarding delay, even if there is no multiplexing delay. ATM Cell Formats The basic ATM cell is made up 53 bytes. The first 5 are header information, and the remaining 48 are for data. In some configurations, 4 out of the 48 bytes may be used optionally by the adaptation layer. Also, the header format at the networknetwork interface (NNI) and the user-network interface (UNI) are slightly different. 7

8 GFC (Generic Flow Control) is used for flow control and fairness to the user-to-network traffic. It is meant for access mechanisms that implement different access levels and priorities. Between two NNI s, it is not used. VPI (Virtual Path Identifier) and VCI (Virtual Circuit Identifier) are used to identify specific network paths and channel for the routing of cells. PT (Payload Type) indicates the type of data (control data or ordinary data) that the cell contains. CLP (Cell Loss Priority) shows the importance attached to the cell. Lower priority cells (those with CLP set to 1) are dropped before higher priority ones) HEC (Header Error Control) is the checksum on the preceding 4 header bytes. This is used to provide some form of error correction and detection. Also, the HEC plays an important role in the final delineation of the cells from the bit stream. ATM Layers The ATM has three layers, namely that ATM adaptation layer, the ATM layer, and the physical layer. These layers correspond roughly to the lower four layers of the OSI - transport, network, data link, and physical layers. The higher layers of the OSI are thus implemented on top of the ATM layers. Physical Layer Transmission and Conversion Sublayer Cell rate decoupling Cell delineation Generation and recovery of transmission frames Adaptation of transmission frames HEC generation and checking Physical medium dependent (PMD) sublayer Bit timing and synchronization Encoding for transmission Transmission (physical medium, electrical/optical issues) ATM Layer Generic flow control Cell header generation/extraction Cell VPI/VCI translation Cell multiplexing and demultiplexing 8

9 ATM Adaptation Layer The AAL adapts the data stream from the higher layers of the OSI to what the ATM can handle, and also packages the data from the ATM layers in a form suitable for the higher layers of the OSI. The general functions of the AAL include: Packet segmentation and re-assembly Handling of cell delay variation (usually by use of a buffer) Handling of lost or miss-aligned cells (by using the sequence count) Bit error checking and correction Convergence Sublayer Performs the adaptation of different types of applications to the ATM. This is necessary for the ATM to support different types of service with different traffic characteristics. The functions performed here are usually service specific. Segmentation and Re-assembly sublayer At the sender, segmentation of higher-level protocol data units (PDU s) into ATM fixed size cells At the receiver: re-assembly of the ATM fixed-size cells into suitable PDU s for the higher levels Call Setup and Call Routing in the ATM The service provided by the ATM is connection-oriented. Thus, a connection must be set-up between the sender and receiver before data cells can be transmitted. This provides the ATM the opportunity to perform two important functions: QoS negotiation & admission test between the end system and the network, and between the intermediate networks Assignment of VPI and VCI to the connection, if the desired QOS requirements can be met The general call setup procedure in the ATM is as follows: 1. Sender makes call request with QoS specification 2. Determination of outgoing network connections 3. QoS negotiation and admission test 4. Allocation of VCI s and VPI s 5. If there are other intermediate switches, perform 2 to 4 for all the intermediate network switches 6. Confirmation of VPI s and VCI s from the target end system (the receiver) to the intermediate network switches, back to the sender 7. The sender now sends the data using the VCI and VPI provided by its nearest switch along the agreed-upon path in its header. The VCI s and VPI s are modified as the cell circumnavigates over the agreed-upon paths. Each ATM switch maintains a routing table of the VCI s and VPI s for each connection that goes though it. The general format of the routing table is as follows: <incoming link, incoming VPI, incoming VCI; outgoing link, outgoing VPI, outgoing VCI > The routing table is updated after the call setup process. Routing is then performed by simply following the entries in the routing tables, and by changing the header entries for the VPI s and VCI s as the cell is passed from one switch to the other. 9

10 ATM Service Classes The ATM provides four service classes. The classes differ in their QoS attributes and the type of traffics they support. The table below shows the service classes, and the different AAL layers used to support the classes. Service Class Class A e.g. CBR video Class B e.g. VBR voice and video Class C e.g. bursty data service Class D e.g. Bursty datagram service Timing Relation Required Required Not Required Not Required between sender and receiver Bit Rate CBR VBR VBR VBR Connection Mode Connection- Oriented Connection- Oriented Connection- Oriented Connectionless AAL Type AAL1 AAL2 AAL3/4, AAL5 AAL3/4, AAL5 ATMs and Multimedia Communication It is generally agreed that of all the currently available networking technologies, the ATM with its huge bandwidth is the most suitable for MM communication. Some of the reasons are listed below: Large bandwidth QoS guarantees Efficient use of network resources Multicasting Scalability Versatility uniform multiplexing and switching technique independent of bit rate, or transmission medium. QoS Metrics in ATM Networks ATM networks are connection-oriented, packet-switched networks. The QoS issue on ATM networks can thus be considered from two aspects: call control parameters for connection-oriented networks and information transfer parameters for packet switched networks. Call control parameters: Connection setup delay ( 4500ms for 64kbps ISDN at the maximum distance of 27,500km) Connection release delay ( 300ms ) Connection acceptance probability (or blocking probability) Information transfer parameters These are QoS parameters requested by applications. Among others, the network uses these to : Determine if a connection can be admitted, Find a path that can support the required QoS Allocate resources The QoS parameters include: BER bit error ratio 10

11 CLR cell loss ratio: ratio of the number of lost cells to the total number of cells transmitted. Can be caused by bit errors or buffer overflow. The CLR is also used to define the average time between cell losses ATBCL. CIR cell insertion ratio: ratio of the number of inserted cells to the total no of cells transmitted. Can be caused by header errors that are not detected by the HEC. Could lead to loss of synchronization, or unexpected traffic increase. CTD end-to-end cell transfer delay CDV cell delay variation Skew an indication of the allowable loss in synchronization between streams. Practically, this is the difference between the presentation times of two related objects Some service attributes for B-ISDN Applications Sevice BER CLR CIR Delay(ms) Telephony with echo cancelers without echo cancelers < 25 <500 Data transmission Distributive computing Hi-fi sound Remote process control Recommended Error Rates for some B-ISDN Applications Application Bit rate BER 1 BER 2 CLR 1 CLR 2 (Mbps) Videophone 2 3x x x10-6 Video conference x10-6 4x10-9 5x10-6 TV distribution x x x10-7 MPEG x x x10-6 MPEG x x10-6 2x10-9 4x without error handling at the AAL 2 with single bit error correction and additional cell loss correction at the AAL 11

12 MM Traffic Characterization Traffic specification Usually, an application needs to provide its traffic specification to the network. The traffic specification defines the minimum set of parameters that the application can be expected to declare while providing the network management as much information as possible for effective control of network traffic and for high network utilization. For the network to know if it can meet the service requirements of an application, it needs further information about the nature of the traffic that it carries. This is provided by a characterization of the traffic. Thus, traffic characterization is needed for two basic reasons: To provide QoS support For optimal utilization of network resources (network bandwidth and buffer space) Active and Silence periods The basic method approach to model traffic sources is by use of active and silence periods sometime called ON/OFF periods. During a connection, a source may go through series of active and silence periods. An active period is the period during which the source generates some data. A silence period is the interval between successive active periods. Thus no data is generated during a silence period. The following are used to characterize multimedia traffic Peak rate Average rate λ p maximum bit rate at which a source can generate data during the active periods λ a number of bits generated over a long period of duration T = length of the duration T Burstiness β measures the amount of variation in the data generation. Traffic burstiness can be described by the following: peak bit rate β = average bit rate Average burst length = average length of the active period during which a source generates traffic at its peak rate Another parameter is the average time between the start of two successive active periods Two basic types of service Constant-Bit Rate (CBR) Service The bit rate is constant over the duration of the transmission. Constant bit-rate can be as a result of the source encoding used, or by use of smoothening buffers. Here, either the source is always active for the entire session, or silent periods are also transmitted at the peak rate. The problem is how to find a constant bit rate that minimizes the required network bandwidth, while maintaining the required QOS. CBR streams can be characterized by using only the peak bit rate, λ p Burstiness, β=1 Variable Bit-Rate (VBR) service VBR could result from either alternation between active and silence periods, or from the transmission of continuous bit streams at varying rates. VBR characterization is more difficult than that of CBR. 12

13 Traffic Models Traffic models try to capture the characteristic of the multimedia traffic by use of various parameters. Most also try to exploit the various forms of correlation in the traffic: Correlation between contents of consecutive data units, example in VBR video Correlation between arrival packet rates Example models that have been proposed include: ON/OFF Models alternates between burst (ON) and silence (OFF) periods. A burst consists of a random number of consecutive data packets. This has been used for voice transmissions. Fluid Models: Source is considered as a stream of fluid characterized by a flow rate. Packets are assumed to be infinitesimally small as compare to the available bandwidth. The source also alternates between ON and OFF periods whose duration are assumed to be iid random variables. During the ON periods, infinitesimally small packets arrive at a constant rate similar to fluid flow. This is an appropriate model for the ATM with its small fixed sized cells, and large bandwidth. Time Series Models This has been used to model data traffic of compressed video streams. This exploits the usually constant inter-arrival times for frames (since play-out rate is fixed), and the high correlation between consecutive frames. The size of the next frame is thus approximated from those of the current and past frames using an auto-regression model. (r,t)-model: Under the rate-interval model, no more than rt bits are generated during any interval of length T. D-BIND -Deterministic Bounding Interval-Dependent Model Here, rather than using one (r,t) pair, a family of rate-interval pairs are used to characterize the traffic. For each pair, the rate is the bounding rate over the interval length. Other models that have been used include Markov and Poisson models. Traffic Shaping Admission control, scheduling, and resource allocation are some of the general methods used to provide QoS guarantees. However, at times users may attempt to exceed the rates specified at the start of the connection, the network thus needs further methods (traffic policing) to ensure that the traffic generated in what it can handle. Multimedia traffic shaping techniques are used to make multimedia traffic amenable to analysis and possible policing. Basically, the source generates data at its peak or average rates, and submits to the shaper. For VBR, the input data will typically be bursty. The shaper outputs (to the network) data at a constant rate, or bursty data that are easier to describe or model. Usually, If If λ λ > λ a λ λ o a λo λp, where o o p, there could be instability in the queue. λ is the output rate from the shaper., there will be low bandwidth utilization Uses To smoothen complex traffic that may be difficult to be support directly Saves bandwidth, especially for highly bursty traffic 13

14 Traffic shaping inherently introduces some delay. The amount of shaping is directly related to the amount of delay introduced. Thus, the timeliness of the data will need to be considered in determining the extent of the shaping that can be applied. Desirable properties of a Traffic Shaper Should be able to shape different traffic types Output should be easy to describe and police Should be easy to implement Simple Leaky-Bucket Incoming data is placed in a buffer of finite size B, (called bucket). The data is then allowed to leak out of the bucket at a constant rate, λ. The value of B thus determines how much data that may be stored in the buffer before onward o transmission to the network. When the bucket is full, incoming cells are droped. B thus determines the burstiness that can be supported from the source. The delay experienced by incoming data is simply given by Clearly, λ a λo λp. o bandwidth of the traffic. λ is usually chosen as the effective bandwidth B λ. The delay increases with increasing B or decreasing λ o. o λ eff, which is typically taken to be the average Since cells are discarded when the buffer is full, or may experience long delays for low output rates, B and carefully chosen. λo must be Advantages of the Simple Leaky Bucket Simple and easy to implement Output traffic is easy to describe and police Since effective bandwidth is additive, by knowing the effective bandwidth for the different input traffics, the network can easily know if it can support an incoming traffic. Problems Dependence on λ eff in most multimedia traffic, it is difficult to determine one unique λeff for the traffic. Requires large buffer space (or it will be applicable for traffic with high data loss tolerance) Large buffers on the other hand imply long delays (r,t)-smooth Shaper (also called Jumping Window Shaper) This acts like the (r,t) traffic model. Within a time duration of length T (called window) a maximum of r cells are allowed to be transmitted. So the worst case buffer requirement will be 2r. This occurs when the maximum r cells are released in two adjacent windows one at the end of the preceding window, and the other at the start of the current window. 14

15 Token Bucket Shaper The token bucket shaper is a generalization of the leaky-bucket. The objective is to handle more bursty traffic without needing large buffer space. The token bucket regulates the transmission of cells by use of tokens. If there are n tokens, n cells can be transmitted. Which means no cell can be transmitted when there is no tokens. It also means that n cells can all be transmitted in one burst. When the token bucket is full, tokens (not data cells) are discarded. The extra cells may still be buffered, or marked for low priority. For stability, λ a λ t λ p The output of the token buffer is no longer at constant rate, but a could be a variable (though bounded) rate. Within an interval of T, the maximum output that can be sent to the network is C + λtt. Long term output smoothers: n( T) λtt. So the long-term output rate is λ t. Overall, The token bucket could cause problems in policing, since there could be bursty traffic. The traffic is however easy to describe, and the burstiness is bounded. Other Shapers Moving window shapers Composite shapers Traffic Shaping and Bandwidth Allocation In general, the bandwidth for a shaped traffic can be allocated by use of the traffic shaping parameters. For Leaky-bucket, the bandwidth can be allocated at a fixed rate of For Token-bucket, the minimum bandwidth allocation will be λ bw λ = λ At the access multiplexer, to support m shaped streams, the output link of the multiplexer can be determined as: λ m i= 1 λ bw( i ) In general, as λ (or λ ) approaches λ we can achieve large statistical multiplexing, which implies better bandwidth o t a bw λ utilization. However, small output rates imply more delays for the cells, and possible cell loss. Thus, there is a tradeoff between the bandwidth utilization that can be achieved by the shaper, and the delays introduced. t o 15

16 Multimedia Traffic Scheduling, Admission Control, and Policing Admission Control Admission control algorithms are used to determine whether a new request for connection is to be granted or refused. A new connection is accepted only if the network can support the QoS requirement of the new connection, without adversely affecting the QoS provided to the currently existing connections. Two Approaches Performance-based QoS satisfaction of the user is most paramount. Economic-based the profitability of the connection to the network provider is given more priority In providing admission control, the following questions will have to be considered How can we know the bandwidth requirement of the new connection? How can we ensure that the multiplexing of this new connection with the existing connections will not affect the other connections? How can we make these decisions in real-time? How can we maximize network utilization bearing all the above in mind? Once again, the characterization of the traffic is required to answer the above questions. Superimposition of multiple arrival streams To determine the requirements of the new connection, one approach that is used is to superimpose the existing connections, and treat them as one connection with known characteristics. Then, the new connection and the superimposed connections are treated as the only inputs to the network switch, based on which we can determine if certain requirements of the new connection can be met. Bandwidth Allocation Bandwidth is usually allocated by use of multiplexing. Two types of multiplexing used are deterministic and statistical multiplexing. Deterministic multiplexing - bandwidth for each connection is allocated using the connections peak rates. If C = the link capacity, λ = bandwidth allocated to the i-th connection, and λ = peak rate of the connection, then: i bw i i n i λ bw = λp and =1 λ bw C i Some of the problems include: Possible waste of resources, especially for bursty traffic Could reduce congestion to a negligible probability, but still cannot eliminate it. Buffer overflows can still occur, depending on the buffer size. Thus cell loss can still be encountered. Does not support efficient sharing of network resources i p Statistical multiplexing Here each connection is allotted a statistical bandwidth which is necessarily greater than the average rate, but less than the peak rate. Thus the sum of the peak rates could exceed the available bandwidth, but the total statistical bandwidth must be no greater than the available bandwidth. That is: λ i a i i n i λbw λp and =1 λ bw C i 16

17 This allows more connections to the network. The assumption is that the different connections would not start sending data at their peak rates at the same time. The performance of statistical multiplexing in terms of the multiplexing gain thus depends on both the characteristic of the new connection, and that of the existing connections. Key issues to consider include: burstiness, burst length, the ratio of the peak rate to the available bandwidth. Admission Control Algorithms Different methods for connection admission control have been proposed. Some consider the actual traffic, while others do not. Also, some consider available resources, such as the buffer size, and the network capacity. Others use a combination. Some of the methods are based on: Gaussian approximation of the superimposed traffic Equivalent capacity methods Approximation of cell loss rates Traffic observation MM Traffic Scheduling Admission control determines which new connections can be admitted to the network without violating the QoS guarantees of existing clients. Once the connection has been admitted, it means that the network should have enough bandwidth to support the new connection. Also, the network must have judged that it can provide the required delay guarantees. To provide the actual guarantees, we still need to consider the order in which the connections are serviced. The order in which the connections are serviced is determined by the scheduling discipline that is adopted. In the network, the major resource being competed for is the network switch, where the connections compete for switch processing time and the output link. Thus, queuing scheduling disciplines are often used to determine the order with which connections will be served at the switch. Traffic Characteristics guaranteed Depending on the scheduling discipline and the multiplexing policy used, guarantees on specific traffic characteristics could be provided at the switch. These are guarantees on the maximum delays experienced by the packets, and guarantees on the minimum bandwidth available to a connection. Delay Guarantees These usually rely on priority-based scheduling schemes. The worst case delay experienced by a packet at a particular switch is bounded. The main issue is in how the priority is assigned. Two approaches have been adopted: Static priority assignment each connection is assigned a static priority at connection set up time. Upon arrival, each packet is stamped with the priority label for the connection, and then added to the common queue. Cells in the common queue are then serviced based on their priority. Dynamic priority assignment packets that belong to the same connection can have different priorities. The priority depends on the state of the server, and the other connections. A typical example here is the EDF (Earliest Deadline First) scheme, in which real or virtual deadlines are assigned to data packets. One problem here is that delay guarantees often require the characterization of the traffic in order to provide delay bounds. This is usually done by traffic shaping schemes. However, the priorities could alter the original shape of the traffic, which means that we may need to reshape the traffic into its original form after it exits from the switch. 17

18 Bandwidth Guarantees If the traffic characterization is known, (for instance the peak and average rates), the switch can provide guarantees on the minimum bandwidth available to each connection. Bandwidth guarantees are often provided by use of some fair-queuing disciplines or frame-based schemes. In delay guarantee schemes, the end-to-end delay bound is dependent on the delays experienced on all the intermediate nodes. For bandwidth guarantees, the end-to-end bandwidth guarantee can be provided by considering only the bottleneck node, i.e. the node that may find it most difficult to provide the minimum required bandwidth. Performance Criteria The suitability of a given queuing discipline can be judged by the following measures: Type of guarantee statistical or deterministic guarantees Connection properties guaranteed upper bounds on delay, upper bounds on delay jitter, upper bounds on error rate (or loss probability), minimum bandwidth Assumptions made Ease of implementation Service Scheduling Disciplines Scheduling disciplines can generally be grouped into two classes: Work conserving the server is not allowed to be idle as far as there is some work to do i.e. some connections waiting to be served. A typical example is the FIFO, whereby a packet is always sent in each time slot whenever the buffer is not empty. Non-work conserving each packet is assigned an eligibility time. The server serves only the packets that are eligible. When no packet is eligible, the server just stays idle. An example here is the jitter-edd and most framebased schemes, such as Stop and Go. This classification has important implications for the buffer requirement, delay and delay jitter that may be experienced at the server. Clearly, non-work conserving schemes will require more buffer space, and could introduce more delays. However they can bound the delay jitter experienced in the server. Work conserving disciplines may not be able to provide tight bounds on delay jitter. Scheduling Disciplines Virtual Clock In VC, each packet is assigned a virtual transmission time. Packets are then transmitted in increasing order of the transmission time. The motivation is to emulate the synchronous TDM, in which connections are allocated periodic transmission time slots. The virtual transmission time is assigned without consideration of the other connections to the server. Rather it is based on the arrival rate at the connection concerned, and the service rate to that connection. 18

19 FQ and WFQ Fair Queue (FQ) is a generalization of the head-of-line processor sharing service discipline. Packets from each connection are placed in a FIFO. Assume we have n non-empty queues, then the server serves each of the queues at a rate of n C, where C is the link capacity. This means that the connections share the link bandwidth equally. Theoretically, this can be is done by doing a bit-round robin for all the connections. Practically, this is achieved by using packets rather than bits. Each packet is assigned a finish number, which corresponds to the time at which the packet would have been served, assuming bit-round robin. Packets are then served in increasing order of their finish numbers. Since different connections may require different bandwidth allocations, it may be wasteful to allocate each connection the same bandwidth. The Weighted Fair Queues solves this problem by assigning weights to each connection. The weight determines the proportion of the bandwidth the connection can have at anytime. The weights are then translated into finish numbers, by calculating the number of bits of service the connection should receive during a round. Here, the service rate to connection i will be: S i = w i n j = 1 w j C where w i is the weight for connection i. Problems with WFQ Difficult to implement Delay bounds could be long Applicability depends on how well the traffic is characterized Delay-Earliest Due Date (Delay-EDD) This is an extension of the EDF scheduling discipline. In EDF, each packet is assigned a deadline, and then packets with the nearest deadlines are transmitted first. In Delay-EDD, a QOS contract is entered between the switch and the connection. The server promises a target deadline which will be met if the connection sends traffic according to the contract agreement. The scheme depends heavily on the traffic characterization (e.g. the peak and average rates), since this is used to estimate the target deadline. The deadline is usually determined using expected packet arrival time (which depends on the arrival rate) and the delay at the server. Jitter-Earliest Due Date (jitter EDD) This is a non-work conserving service discipline. It is an extension of the Delay-EDD to provide jitter bounds, in addition to delay guarantees. Each packet that is served at a server is given a time stamp indicating the difference between its deadline and the actual time that service was completed. At the next switch, the packet is delayed for this period of time before it is scheduled. Thus, the packet could experience the maximum delay at each node along its path. However, the delay variation is incurred only at the last switch. Thus, like in Delay-EDD, delay is bounded (by the sum of the delays in each node). Unlike delay-edd, the jitter is also bounded by the jitter at the last one node. Stop and Go Here, time is divided into fixed length intervals called frames. The arriving frame of each input connection is mapped to the departing frame of the output link, by introducing a constant delay (< T). Packets that arrive at the server at frame f are thus sent out only at the next frame, f +1. Thus, at each node, a packet can stay in the server for at most 2T time units. Stop and Go guarantees that packets transmitted at the same time frame from the source stay in the same time frame throughout the network. Thus, the maximum delay jitter is the duration of the frame, T. One Other interesting scheduling discipline is the generalized round robin. 19

20 MM Traffic Policing Admission control and traffic scheduling may not be sufficient to ensure QoS guarantees. A user may sometimes use more resources than those previously negotiated (and allocated) due to a number of reasons: User stated requirements may be different from the actual requirements e.g. a user may not understand the actual requirements, or may knowingly underestimate the requirements (for instance, for cost reasons) Faults or problems in the network Some users may decide to be malicious The network must thus monitor (police) the traffic from each of the sources and take appropriate actions to ensure that the traffic stays within its negotiated parameters. Traffic policing should be done in such a way that currently conforming users do not experience QoS degradation. This means that non-conforming traffic should be detected as soon as possible. Traffic policing mechanisms usually depend on the traffic characterization used. For streams that have undergone some form of traffic shaping, their output behaviour is easier to characterize, and hence easier to police. For more complex characterization, the resulting traffic will be more difficult to police. Thus, the traffic parameters to be monitored and controlled will be those used to characterize the source traffic. The parameters of interest here include: Actions that may be taken against violating traffic λ, λ and the length of the active period. i a i p Treat conforming traffic with higher priority, and mark violating traffic as low priority Delay violating traffic in a buffer, such that the output from the buffer conform to the negotiated contract Adaptive traffic control by informing the sender when violation is detected Discard violating traffic Traffic shaping methods can also be used to take actions against violating traffic - for instance, when the traffic may need to be delayed (or shaped) to the negotiated traffic pattern. Selective Discarding Traffic policing aims at protecting the network form excessive traffic, and meeting the QoS guarantees of existing users. Explicit discarding of violating data packets may not be the best approach, since users may not even know that they are violating the contract. Moreover, the problem may be due to other problems resulting from the statistical nature of communication networks. Discarding is thus usually done only when the offending traffic is likely to lead to QoS violations for other users. Two approaches that have been proposed are: Push-Out non-conforming cells are marked as low priority, while others are marked as high priority. When switch buffers are full, high priority cells push-out low priority cells in the buffer. Thus, low priority cells are allowed only after higher priority cells have been accepted, and there is still space. The problems here are: Loss of cell sequence Complexity and overheads queues are no longer FIFO, but priority-based Thresholds A fraction of the available buffer space is left for exclusive use by high-priority packets. The remaining proportion is shared by high priority and low priority cells on FCFS basis. Determining the appropriate thresholds could pose problems. This usually will require a characterization of the traffic. 20

21 Dynamic Source Coding When the network notifies a source of traffic violation, what can the source do? We expect it to reduce its traffic, or to stop traffic temporarily. But what is actually done, and how it is done could depend on the type of traffic. For instance, for real-time traffic, the source may not be able to stop transmission. The general approach will depend on if the traffic is delay-insensitive simply delay traffic by buffering delay-sensitive buffering may be difficult due to the need for large buffer space. The options may be to reduce traffic rate, mark cells with priority, increase quantization step size for the analogue to digital conversion. The issues to consider here include the propagation delay, and the network bandwidth. In fact, the constraint is on the "bandwidth- propagation delay product". This product determines the amount of data that might have been released into the network before violation notice is received at the source. Thus, the effectiveness of adaptive traffic policing schemes depends on this factor. You can imagine the case of an ATM with huge bandwidth, over a wide area, when we have already pumped many cells into the network! Overall, we can group the policing/control schemes as Proactive prevention, which is better for guaranteed service Reactive prefers to cure, which is more appropriate for best effort service 21