Method, an output unit, a terminal, and computer program products for reconstructing non-continuous packetized data of a continuous data stream from a data connection into a continuous data stream at the receiving point of a packet-based network

Abstract

This invention relates to a method for reconstructing non-continuous packetized data of a continuous data stream like streamed media, voice, audio, or video from a data connection into a continuous data stream at the receiving point of a packet-based network, comprising the steps of

Description

FIELD OF THE INVENTION

[0001] This invention relates to a method for reconstructing non-continuous packetized data of a continuous data stream like streamed media, voice, audio, or video from a data connection into a continuous data stream at the receiving point of a packet-based network as well as, an output unit, a terminal, and computer program products therefore.

[0002] The invention is based on a priority application No. 02 360 111.5, which is hereby incorporated by reference.

BACKGROUND OF THE INVENTION

[0003] Many sophisticated emerging applications, such as voice over IP, multimedia conferencing, or distributed virtual reality, are difficult to deploy in todays internetworking infrastructure. This is mainly due to one requirement that all these applications share the need for guaranteed real-time service. These applications not only require high bandwidth, but predictable quality of service (QoS) such as jitter delay as well.

[0004] The QoS requirements at network level are typically specified in terms of bounds on worst-case end-to-end delay on the worst-case packet loss rate and on the worst-case delay jitter for packets of the connection. Other parameters may be specified as well, such as deadline miss rate. The desired delivery time for each message across the network is bounded by a deadline, a specific maximum delivery delay. This delay bound is an application-layer, end-to-end timing constraint.

[0005] If a message arrives after the deadline is expired, the message is useless and is typically discarded. For many real-time applications, it is not important how fast a message is delivered. Indeed, packets arriving early may need to be buffered at the receiver to achieve

[0006] constant end-to-end delay. Therefore, delay jitter, which is the variation in delay experienced by packets in a single connection, is a critical performance metric. For example, in video transmission, jitter may cause some frames to arrive early, and others to arrive late. Although the transmission of all frames satisfies the deadline requirement, the displayed movie may appear jittery. Same applies to streamed audio data.

[0007] Buffers at the receiver can be used to control delay jitter. The amount of buffer space required can be determined from the peak rate and the delay jitter of the delivery process and can be quite large for a network with no control of delay.

[0008] Important quality of services are especially delay jitter, delay, and packet loss. Delay jitter and packet loss obstructs proper reconstruction at the receiver whereas delay impairs interactivity.

[0009] The following section contains definition for the notions of streams, packets, and channels.

[0010] Streamed data is a data sequence that is transmitted and processed continuously. Streaming is the process of continuously appending data to a data stream.

[0011] A packet is a piece of data consisting of a header and a payload information. Packetizing is the process of decomposing data into a set of (small) packets, where the header is used to store information for reconstruction, e.g. a sequence number.

[0012] A data channel is a connection between two network units that is able to transport data.

[0013] Delay is the time between sending and receiving a packet. Delay jitter is the variation in delay. Packet loss is an infinite delay.

[0014] A common, used technique for streamed data is to use a buffer at the receiver for reducing delay jitter and packet loss against an increased overall delay. Hence there is a demand for optimization. Especially real-time streamed data, like video or audio streams, needs to be on-line processed, i.e., with small delay and small jitter delay.

[0015] A well known algorithm to solve this problem is to buffer streamed data and to replay the buffer at a constant speed to absorb delay variations and play-out packets at fixed deadline, called jitter absorption. Packets received after deadline are discarded.

[0016] A more sophisticated algorithm is to monitor delay and/or delay variation and adapt play-out time accordingly, called jitter adaptation. An application might then slow down play-out when delay increases to avoid loss and speed up play-out when delay decreases to reduce delay.

[0017] It is object of the invention to provide a method for reducing delay jitter, delay, and packet loss for streamed data connections.

SUMMARY OF THE INVENTION

[0018] The present invention is a method for reconstructing non-continuous packetized data of a continuous data stream like streamed media, voice, audio, or video from a data connection into a continuous data stream at the receiving point of a packet-based network, comprising the steps of

[0019] providing of at least one estimation method based on at least one characteristic value concerning data connections of the kind intended for,

[0020] gathering measurements of at least one value characterizing the data connection,

[0021] evaluating a de-jittering delay for the data connection by predicted parameters taking into account the at least one provided value and the at least one gathered value,

[0022] delaying and converting the data packets following the evaluated de-jittering delay.

[0023] The invention also relates to an output unit, a terminal, and computer program products for a terminal and for an output unit.

[0024] The essential idea of the invention is iterative gathering network observations for a statistical prediction of network behavior, and adapting iterative said converting method according to said prediction. The present invention uses a continuous optimization for adapting the parameters of a conversion method. This optimization decomposes into three steps. Continuously gathering network observations, i.e. quality of service measurements, deriving a statistical prediction from these network observations, and adapting the parameters of the conversion method according to said prediction.

BRIEF DESCRIPTION OF THE DRAWINGS

[0025] FIG. 1. shows a network, terminals, an output unit, and the context of streamed data reconstruction according to the invention.

[0026] FIG. 2. shows the phases of reconstructing streamed data out of a packet stream according to the invention.

[0027] FIG. 3. shows a use case diagram according to the UML notation describing the boundaries of the method for streamed data reconstruction according to the invention.

[0028] FIG. 4. shows a class diagram according to the UML notation describing an architecture of computer program for streamed data reconstruction according to the invention.

[0029] FIG. 5. shows a computer program for streamed data reconstruction according to the invention.

[0030] FIG. 6. shows an “Estimation” class for a computer program for streamed data reconstruction according to the invention.

[0031] FIG. 7. shows a timeline of a encoding—transportation—decoding delay scenario.

DETAILED DESCRIPTION OF THE INVENTION

[0032] FIG. 1 shows of a network B1 two data channels B2, an output unit B3, and two terminals, a computer terminal B4 and a telephone terminal B5. The terminal B4 has an output unit B3. This output unit B3 is connected via a data channel B2 with a network B1. The telephone terminal B5 is as well connected with the network B1 via a data channel B2.

[0033] The figure describes the scenario for this realization. Both terminals B4, B5, in the role of a receiver, are connected with the network B1 via data channels B2. The terminals receive packets over the data channels and these packets contain streamed data, which has to be reconstructed. To be able to reconstruct the data stream, there might be a special hardware, called output unit B3, that alternatively might be integrated in the terminal. The terminal and the output unit are assumed to be controlled by a computer program. Although the realization of the reconstruction method could also be implemented in software only.

[0034] FIG. 2 shows a control entity A1, a buffer queue A2, an input channel A3, an output stream A4, an input packet sequence A5, an output data stream A6 and an illustration of two time intervals A7 between two consecutive packets also-known as packet inter-arrival times.

[0035] The control entity A1 controls the buffer queue A2, i.e. when the queue has to be emptied and filled. The buffer queue A2 is connected with the input channel A3 transporting the input packet sequence A5. The input packet sequence A5 consists of a sequence of packets A5, where each packet having a packet sequence number 15, 16, . . . , 20. This input packet sequence AS needs not coinciding with the packet number sequence as illustrated in the drawing. The figure does not show the packet representation, i.e. header, payload, etc. It is assumed that the payload is already extracted and labeled by the sequence number. The figure shows especially the time intervals A7 between the consecutive packets 19 and 20 as well as the time intervals A7 between the consecutive packets 15 and 16. The buffer queue A2 is also connected with the output stream A4 transporting the ordered continuous output data stream A6. The output stream is ordered by packet numbers and the time interval between two consecutive packets disappears, by using the previously buffered reservoir.

[0036] In the illustrated configuration the output stream data carries data from packets 1, 2, 3, 4, 5, the buffer queue A2 stores packets 6, 7, 8, 9, 10, 11, 12, 13, and the input channel data AS consists of the packets 15, 14, 16, 17, 19, 18, 20.

[0037] The figure illustrates the functionality of reconstructing a data stream. A jittered input data stream running into a buffer, converted into a continuous output data stream. The arriving packets, each having its number, are translated into an ordered continuous data stream where the data is ordered by the packet numbers and the time interval between the content of two consecutive packets disappears. In the example it is assumed that the packet stream has a jitter and the packets need not arrive in the origin sequence. The network might have additional characteristics, e.g. an asserted delay bound that should be taken into account when implementing the described functionality. In further, it is assumed that there is no packet loss. In case of packet loss additional strategies have to be considered beside buffering, e.g., reconstruction of packet information on the application layer or depending if network resources and time are available an additional request for retransmission.

[0038] FIG. 3 shows a use case diagram according to the UML notation, from the ‘Unified Modeling Language User Guide’, G. Booch, J. Rumbaugh, I. Jacobson, Addison-Wesley, Reading Mass., 1999, pages 233-236, containing the actors “Network” and “Application”, as well as a use case “Converter” and a use case “Control”.

[0039] The “Network” is associated with the “Converter” by “Data channel” and the “Application” is associated with the “Converter” by “Data stream”. The “Converter” is extended by the “Control”.

[0040] The diagram shows the problem context, namely the data channel “Data channel” supporting the jittered packet data stream shown in FIG. 2, and a application “Application” requesting the reconstructed continuous streamed data. This reconstruction is performed by a controlled converter “Converter” extended by “Control”. The control mechanism is explicitly stated. It might be hidden by other use cases as side effects, e.g. a scheduler integrated in an operating system.

[0041] FIG. 4 shows a class diagram according to the UML Notation, from the ‘Unified Modeling Language User Guide’, G. Booch, J. Rumbaugh, I. Jacobson, Addison-Wesley, Reading Mass., 1999, pages 105-108, containing the data types “Channel”, “Stream”, and “PriorityQueue”; the processes “Receive” and “Stream”; and a class “Estimation”.

[0042] “Channel” provides the two methods “End” and “Fetch”. “Stream” provides the two methods “Append” and “Read”. “PriorityQueue” provides four methods “Add”, “Get”, “IsEmpty”, and “Size”. “Estimation” provides the two methods “Measure” and “Predict”.

[0043] The diagram shows an architecture for streamed data reconstruction. This architecture has a framework character. It is designed for illustration purposes. It allows to substitute the estimation and to simplify the description by abstraction. An architecture of a realization is influenced by the complete product design.

[0044] The architecture consists of three abstract data types, a channel, a stream and a priority queue, as well as two processes, “Receive” and “Stream”. The priority queue is chosen to illustrate the abstract buffering mechanism. It is not necessary to use abstract data types. For instance, a often used technique instead of a priority queue is a straight forward array implementation of a buffer queue.

[0045] The processes need not to be explicitly designed. Instead one might realize the method by threads or operating system services.

[0046] The data type “Channel” is aggregated by the process “Receive”. The data type “Stream” is aggregated by the process “Stream”. The data type “PriorityQueue” and the class “Estimation” are both associated to both processes “Receive” and “Stream”.

[0047] The method “End” of the data type “Channel” returns the Boolean true when the last packet of the packet sequence has arrived, the Boolean false otherwise. The method “Fetch” returns the next received packet.

[0048] The method “Append” of the data type “Stream” appends the argument to the end of this stream. The method “Read” reads the head of this stream (destructive).

[0049] The method “Add” of the data type “PriorityQueue” enters the argument into this priority queue.

[0050] The method “Get” returns the least element of this priority queue. The method “isEmpty” returns the Boolean true if this priority queue contains no element, the Boolean false otherwise. The method “Size” returns the number of elements contained in this priority queue.

[0051] The method “Measure” of the class “Estimation” collects network performance information and updates network characteristics accordingly. The method “Predict” returns values for controlling the behavior of the two processes. The two processes are controlled by the class “Estimation” that measures network behavior and derives network performance predictions. The two processes “Receive” and “Stream” use this prediction in order to adapt their behavior, e.g. the use of the buffer queue or the stream speed etc.

[0052] FIG. 5 shows a program implementing the architecture for streamed data reconstruction of FIG. 4.

[0053] The abstract notation for the program consists of a declaration part for variables and types, labeled by ‘DECLARATION’ and an implementation part labeled by ‘IMPLEMENTATION’.

[0054] The variable declaration part consists of three objects:

[0055] “Input”, a “Channel”,

[0056] “Output”, a “Stream”, and

[0057] “Buffer”, a “PriorityQueue”.

[0058] The type declaration part consists of three data types:

[0059] a data type “Channel”, framed by ‘DATA TYPE Channel’ and ‘END DATA TYPE Channel’,

[0060] a data type “Stream”, framed by ‘DATA TYPE Stream’ and ‘END DATA TYPE Stream’,

[0061] a data type “PriorityQueue”, framed by ‘DATA TYPE PriorityQueue’ and ‘END DATA TYPE PriorityQueue’.

[0062] The implementation part consists of

[0063] a process “Receive”, framed by ‘PROCESS Receive’ and ‘END PROCESS Receive’, and

[0064] a process “Stream” framed by ‘PROCESS Stream’ and ‘END PROCESS Stream’,

[0065] a class “Estimation”, framed by ‘CLASS Estimation’ and ‘END CLASS Estimation’.

[0066] The data type “Channel” consists of

[0067] a method “End”, returning the Boolean true if the input packet sequence ends, and

[0068] a method “Fetch”, returning the next arrived packet.

[0069] The data type Stream has also two methods:

[0070] a method “Append”, adding a data element at the end of this stream, and

[0071] a method “Read”, returning the next element of the stream.

[0072] The data type “PriorityQueue” has four methods:

[0073] a method “Add”, adding a packet to this priority queue

[0074] a method “Get”, returning and removing the packet with the least element, i.e. the-packet with the least number, from this priority queue,

[0075] a method “IsEmpty”, returning the Boolean true if the priority queue contains no packet,

[0076] a method “Size”, returning an integer, the number of packets contained in this priority queue.

[0077] The process “Receive” consists of a loop, framed by ‘WHILE’ and ‘END WHILE’, with the terminating condition ‘NOT Input.End( )’, and a body consisting of the statement sequence ‘packet=Input.Fetch( )’; ‘Estimation.Measure(packet)’; ‘Buffer.Add(packet)’.

[0078] Hence, the process iterative reads a packet from the input channel, update the performance statistic of the network and buffers the packet, until the last packet is arrived.

[0079] The process “Stream” consists of a main loop, framed by ‘WHILE’ and ‘END WHILE’, with the terminating condition ‘NOT (Input.End( ) AND Buffer.isEmpty( ))’ and a body consisting of the statement ‘Estimation.Predict(BufferSize, DelayTime)’ followed by a sequence of further while loops.

[0080] The first while loop, framed by ‘WHILE’ and ‘WAIT END WHILE’ has the terminating condition ‘Buffer.Size( )<BufferSize’ waits until the buffer is filled according to the predicted value Buffer.Size.

[0081] The second while loop, framed by ‘WHILE’ and ‘END WHILE’, with the terminating condition ‘NOT Buffer.isEmpty( )’ and a body consisting of the statement sequence ‘Output.Append(Buffer.Get( ))’; ‘DELAY(DelayTime)’, empties the buffer and serves the stream continuously with a homogenous by the Estimation predicted delay.

[0082] The latter two loops are iterated until the complete stream is reconstructed.

[0083] The kernel of the described program and the control of the processes and the buffer is the class “Estimation”. This class contains the variable “meanDelay”. In general this class contains variables for measured network characteristics.

[0084] Furthermore, the class “Estimation” consists of a set of variables for the statistical observations and two methods,

[0085] a method “Measure” that updates the network characteristics by observed events, here a packet arrival, and

[0086] a method “Predict”, that returns parameters for the conversion, here

[0087] BufferSize and DelayTime, based on gathered network characteristics. It should be noted that the methods of the two processes are only a specific option model. Beside the stated mode there might be a streaming handshake, forcing faster streams, or an application that might allow a homogenous delay or a smooth increasing delay.

[0088] FIG. 6 shows a program implementing a class Estimation introduced in FIG. 5.

[0089] The class “Estimation” is framed by ‘CLASS Estimation’ and ‘END CLASS Estimation’ and contains five variables, three reals “T”, “sr”, and “tr”, as well as two integers “R” and “n”, and two methods.

[0090] A method “Measure” that updates the mean delay T by an observed packet delay t, as well as the decrement of the number of remaining packets R and

[0091] A method “Predict”, that returns parameters for the conversion, buffer size B and delay time (the reciprocal of the sample rate), based on gathered network characteristics.

[0092] FIG. 7 shows three diagrams, labeled by O1, O2, and O3. The x-axis of each diagram is the time and the y-axis are packets. Diagram O1 shows encoding and packetisation, diagram O2 shows transportation through a network, and diagram O3 shows the stream resuming at the receiver. The figure depicts an encoding-transmission-decoding scenario. There are three observation points O1 at the sender, O2 at the network, and O3 at the receiver.

[0093] Diagram O1 consists of a packet P(1,1) and two occurrences of packet P(2,1). Diagram O2 consists of a waiting packet W(2,1) and two total service time intervals Nstag TS for each packet. Diagram 03 consists of a de-jittering delay Tjit and a decoding delay Tdec.

[0094] The diagrams are connected via three dashed arrows showing a path of packet P(2,1).

[0095] The horizontal double arrows A2 shows a time interval until packet P(2,1) arrives. The horizontal arrow W2,1 shows a waiting time interval of packet P(2,1). A horizontal arrow Nstag TS shows a service time interval of P(2,1), and a horizontal arrow d2,1 shows a delay of packet P(2,1).

[0096] Assumptions for the shown scenario are identical encoding (e.g. Voice Activity Detection or not) and packetisation of the arriving calls, with no time stamps and available packet sequence numbers. Negative-exponentially distributed connection inter-arrival time A2 is assumed at the encoder. Shown in diagram O2 a packet-based network delays discontinuously packets with a deterministic service time Nstag TS. No priorities, no retransmission, no overtaking, no change in routing, only real-time traffic, and no disturbing data traffic is assumed.

[0097] The packet P(2,1) is traced through the described scenario. At the sender this packet is created after the time A2 starting from the creation event of the preceding packet P(2,1). When the first packet is processed the packet P(2,1) enters the network. There it waits for the time W2,1. When the waiting time is passed the network transports the packet within time Nstag TS to the receiver. At the receiver it is buffered for a time Tjit and decoded within a time Tdec.

[0098] The following section contains an example application for a stream transmission scenario where a size of a file to stream is known and a network that delays equally sized packets equally. Then considering the following intermediate scenario enabling one to determine the optimal buffer size for continuos streaming, i.e., the following three events coincide: buffer is empty, the file is completely transmitted, and the buffer is completely streamed. Because of the deterministic delay assumption there is no need for prediction. But the example shows the dependence of the scenario parameters and illustrates the adaptive buffer functionality.

[0099] In an intermediate scenario there is a rest of the stream to transmit at the sender, called rest, of size R, a buffered stream, called buffer, of size B and a played stream at the sender. The above three events coincide when the transmission time for the rest and the time for streaming the rest and buffer is equal. The transmission rate tr is 1/T, the stream rate is a constant, say sr. Then the transmission time for the rest is R/tr and the time for streaming the rest and buffer is (R+H)/sr. Derived from the equation R/tr=(R+B)/sr one concludes the optimal buffer size B=sr/tr*R−R.

[0100] For most packet networks the assumption that each packet is delayed equally is wrong. But one could approximate the real delay with the mean delay of the already transmitted packets instead. The mean delay T(n) for n transmitted packets each having its own delay ti is the sum delay t1+t2+. . . +tn divided by n. For calculation T(n+1) consider T(n+1)=(t1+t2+. . . +tn+tn+1)/(n+1)=((t1+t2+. . . +tn)+tn+1)/(n+1), but (t1+t2+. . . +tn)=n*T(n). Hence T(n+1)=(n*T(n)+tn+1)/(n+1).

[0101] The above discussion is illustrated as an implementation of class ‘Estimation’ shown in FIG. 6.

[0102] The statistical model can be enhanced by observable properties of the network like packet routing, traffic, or network topology, and of the stream content itself, like length pauses and talk spurts in the case of voice data streams, as well as past transmissions or even past connections. The following section describes a more complex application for the special case of reducing delay jitter for a packetized voice network, with minimal delay, i.e., small queues in the context and with the assumptions of FIG. 6. A set of recursive measurement and prediction equations, based on multiple probabilistic models is developed illustrating the claimed method. The main assumptions are a constant inter-arrival time for the packets at the network during active voice, but no constant inter-departure time when arriving at the receiver.

[0103] For this application additionally a probability function which describes the network packet delay behaviour is missing. The delay of the first arriving packet (reference packet) dref is unknown, as well as the sender clock is unknown and the time stamps are unavailable. The application has the property to be able re-configuring the queue while silence phases. Hence this application is an example for a tight coupling of the application layer consuming the transmitted stream.

[0104] For the detailed description the following notations are used for the encoding and packetisation delay factors

[0105] frame size TF

[0106] encoder processing time Tenc

[0107] look ahead TLA

[0108] NF code words per packet

[0109] inter-packet time NF TF

[0110] decoder processing delay Tdec

[0111] and for the network delay factors

[0112] inter-packet time NFTF

[0113] service time per node and packet TS

[0114] number of traversed nodes Nstag

[0115] total service time NstagTS

[0116] statistical waiting time WN

[0117] For the end-to-end delay we say the delay introduced by encoder, packetizer and decoder: Tenc,P,dec=NFTF+TLA+Tenc+Tdec, for the delay in the packet-based network: D=NstagTS+WN, and for the dejittering delay: Tjit.

[0118] The initial values for the statistical model are

[0119] the maximum end-to-end delay is dE2E,

[0120] the number of traversed nodes Nstag,

[0121] the service time per node TS,

[0122] the mean number of created packets per call is {overscore (x)} calculated out of the mean call holding time

[0123] calls per second (dependent on daytime)

[0124] packet frame length TF

[0125] number of packets per frame NF

[0126] The assumed/pre-defined statistical values are

[0127] Number of overall competing connections NIP

[0128] Number of route busy periods M

[0129] Number of competing connections per busy period nm

[0130] The following section contains notations used for the described packet delay calculations.

[0131] Amount of packets from calls arriving after the observed connection i until network arrival instant of packet number r. xkm+i,rmin(pr).

[0132] Number of additional packet arrivals of previous connections between lth connection arrival instant and network arrival instant of packet r from connection i: xkm+i,rmin.

[0133] Probability of j Poisson arrivals during packet producing time interval of a single connection: 1$p_{j,r}=\frac{{\left(\lambda \left(r-1\right)N_F{T}_F\right)}^j}{j!}{e}^{-\lambda \left(r-1\right)N_F{T}_F}.$

[0134] The following section contains an itemization of the used notations for mean delay calculations

[0135] Mean delay of an arbitrary packet: {overscore (d)}(Nstag,TS, {overscore (x)},λ,NFTF)

[0136] Mean absolute relative delay of an arbitrary packet: {overscore (Δd)}(Nstag,TS, {overscore (x)},λ,NFTF)

[0137] Mean delay of the rth packet {overscore (dr)}(Nstag,TS, {overscore (x)},λ,NFTF)

[0138] Average number of cumulative network packet arrivals at network arrival instant of packet number r. {overscore (qr)}({overscore (x)},NIP, {overscore (xrmin)}, {overscore (xrmin(pr))}) and of an arbitrary packet: {overscore (q)}({overscore (x)},NIP, {overscore (xmin)}, {overscore (xmin(p))}).

[0139] Average relative number of cumulative network packet arrivals at network arrival instant of packet number r. {overscore (Δqr)}({overscore (xmin(pr))}) and of an arbitrary packet: {overscore (Δq)}({overscore (x)}, {overscore (xmin(p))}).

[0140] Average minimum amount of additional packets from previous connections at network arrival time instant of packet number r. {overscore (xrmin)}({overscore (x)},λ,NFTF) and of an arbitrary packet: {overscore (xmin)}({overscore (x)},λ,NFTF).

[0141] Average minimum amount of additional packets from calls arriving after the observed connection until network arrival instant of packet number r. {overscore (xrmin(pr))}({overscore (x)},λ,NFTF) and an arbitrary network packet arrival instant: {overscore (xmin(p))}({overscore (x)},λ,NFTF).

[0142] Mean total inter-arrival time of an arbitrary packet: {overscore (I)}(λ, {overscore (x)},NFTF) the Ith call: {overscore (Ii−l)}(λ, {overscore (x)},NFTF), and the rth packet: {overscore (Ir)}(λ,NFTF).

[0143] Mean value of NIP Erlang-(i−l) distributed time intervals: {overscore (Y)}(λ)

[0144] Mean values of the relative absolute total inter-arrival time of an arbitrary packet: {overscore (ΔI)}(λ, {overscore (x)},NFTF) the lth call: {overscore (ΔIi−l)}(λ, {overscore (x)},NFTF), and the rth packet: {overscore (ΔIr)}(λ,NFTF).

[0145] The following list contains the set of values for initialisation and adaptation.

[0146] Packet.arrival instants at the decoder: tDr

[0147] Delay of the rth packet: dr

[0148] Reference packet number ref, which is the number of the first arriving packet

[0149] QoS dejittering delay: Tjit

[0150] Packet loss probability: Ploss

[0151] Maximum allowed end-to-end delay: dE2E

[0152] Number of packets per active voice period xkm+i

[0153] Number of packet losses xloss

[0154] Number of overlong delays xE2E

[0155] Coefficient of variation c

[0156] Hypo-exponential Process FD(t; t1,t2) with mean values t1 and t2.

[0157] Hyper-exponential Process FD(t, p, t1,t2) with the mean values t1,2 and probability p.

[0158] We have two qualities of service bounds, the packet loss restriction Pr└d>dmin+Tjit┘<Ploss, and the delay restriction dmax+Tjit<dE2E.

[0159] The problem of serving continuous streamed voice data is solved by gathering the decoder packet arrival instants tDref and tDr; then approximating the delay of the first arriving packet dref with a pre-calculated mean delay value and calculating the delay of the rth packet out of dr=tDr−tDref+{overscore (d)}−(r−ref)·NFTF, and creating a substitute delay probability function to calculate the maximum tolerated packet delay and consequently the dejittering delay.

[0160] Packets missing the quality of service restrictions for packet loss dr≦tDr−tDref+{overscore (d)}−(r−ref)·NFTF, or equivalently tDr≦tDref+Tjit+(r−ref)·NFTF and the end-to-end delay dr+Tjit<dE2E are discarded.

[0161] The following section contains the variables needed for packet delay calculations.

[0162] The delay of the rth packet produced from the lth connection during busy period m is denoted as dkm+i,r.

[0163] Wkm+i,r denotes the waiting time of packet number km+i,r.

[0164] Ii−l,r describes the total inter-arrival period from the begin of route busy period m until network arrival instant of the rth packet of the lth connection. The total number of network packet arrivals from the beginning of the busy period m until service beginning of the observed packet is named qkm+i,ri−1+r−1++xkm+i,rmin(pr)+xkm+i,rmin.

[0165] Y1−1 is the Erlang distributed time interval of i−1 negative-exponentially distributed successive call inter-arrival time intervals.

[0166] ΔIi−l,r denotes the relative total inter-arrival time of the rth packet produced from the lth call.

[0167] The negative-exponentially distributed encoder inter-arrival time of the lth connection is named Akm+l.

[0168] The following section contains a description sample jitter delay algorithm for voice data streams.

[0169] This prediction is based on gathered the decoder packet arrival instants tDref and tDr; via an approximated delay of the first arriving packet dref with a pre-calculated mean delay value and calculate the delay of the rth packet out of dr=tDr−tDref+{overscore (d)}−(r−ref)·NFTF; and a substitute delay probability function to calculate the maximum tolerated packet delay and consequently the dejittering delay.

[0170] There are two quality of service bounds considered, namely, the packet loss restriction Pr└d>dmin+Tjit┘<Ploss and the delay restriction dmax+Tjit<dE2E.

[0171] The “Measure” method for this example initializes the statistic observations by gathering the following values during call set-up

[0172] the maximum end-to-end delay dE2E

[0173] the highest tolerated probability for packet loss due to jitter problems Ploss

[0174] the number of traversed nodes Nstag

[0175] the service time per node TS

[0176] the mean number of created packets per call {overscore (x)} calculated out of the mean call holding time

[0177] calls per second (dependent on daytime)

[0178] packet frame length TF

[0179] number of packets per frame NF

[0180] for

[0181] the (initial) service time NstagTS

[0182] the packet length NFTF

[0183] the initial mean delay of an arbitrary packet {overscore (d(0))}:={overscore (d)}(Nstag,TS, {overscore (x)},λ,NFTF)

[0184] the initial mean absolute relative delay of an arbitrary packet {overscore (Δd(0))}:={overscore (Δd)}(Nstag,TS, {overscore (x)},λ,NFTF)

[0185] the initial coefficient of variation 2$c^{\left(0\right)}=\frac{\stackrel{\_}{\Delta d^{\left(0\right)}}}{\stackrel{\_}{d^{\left(0\right)}}}$

[0186] to determine the initial delay probability function.

[0187] While the call is active the “Measure” method gathers the packet arrival instants tDr. Then the delay of the rth packet by dr=tDr−tDref+{overscore (d(0))}−(r−ref)·NFTF is calculated. The quality of service restriction for streamed voice data are for packet loss requirement tDr≦tDref+Tjit(0)+(r−ref) NFTF and for delay requirement dr+Tjit(0)<dE2E. For the shown statistical description it is necessary to count the number packets per active voice period xkm+i, packet losses xloss, and overlong delays xE2E.

[0188] The route length Nstag and the service time NstagTS as well as the mean delay 3$\stackrel{\_}{d^{\left(q\right)}}\text{:}=\stackrel{\_}{d^{\left(q-1\right)}}+\frac{1}{x_{k_m+i}}\sum _{r=2}^{x_{k_m+i}}{t}_{D_r}-t_{D_{r-1}}-N_F{T}_F,$

[0189] and the mean value of the relative absolute delay 4$\stackrel{\_}{\Delta d^{\left(q\right)}}\text{:}=\stackrel{\_}{\Delta d^{\left(q-1\right)}}+\frac{1}{x_{k_m+i}}\sum _{r=2}^{x_{k_m+i}}\&LeftBracketingBar;t_{k_m+i,r}-t_{k_m+i,\mathrm{ref}}-\left(r-\mathrm{ref}\right)N_F{T}_F\&RightBracketingBar;,$

[0190] and the coefficient of variation 5$c^{\left(q\right)}=\frac{\stackrel{\_}{\Delta d^{\left(q\right)}}}{\stackrel{\_}{d^{\left(q\right)}}}$

[0191] s updated during a talk spurt.

[0192] In “Prediction” method one calculate dmax(q) choosing the Hypo-exponential probability FD(t; t1(q),t2(q)) function when 0≦c(q)≦1, where t1(q)={overscore (d(q))}·(1−c(q)) and t2(q)={overscore (d(q))}·c(q). And calculate dmax(q) from probability function with respect to packet loss probability out of dmax(q)=FD−1(1−Ploss; t1(q),t2(q)) If c(q)>1 choose the Hyper-exponential probability function FD(t; p(q),t1(q),t2(q)), where 6$t_{1,2}^{\left(q\right)}=\stackrel{\_}{d^{\left(q\right)}}·{\left(1\pm \sqrt{\frac{{\left(c^{\left(q\right)}\right)}^2-1}{{\left(c^{\left(q\right)}\right)}^2+1}}\right)}^{-1}$

[0193] and p(q)={overscore (d(q))}/2·t1(q). Calculate the maximum relative delay dmax(q) out of the Hyper-exponential probability density function with e.g. the decomposition method.

[0194] The result is used to adapt the stream output respectively by the maximum relative delay:

[0195] Δdmax(q):=dmax(q)−dmin=dmax(q)−NstagTS and determine Tjit(q) according to Δdmax(q)=:Tjit(q)≦dE2E−dmax(q) during a silence period.

[0196] The delay of the rth packet of the lth connection during busy period m is the sum of its service time and its waiting time in the network: dkm+i,r=NstagTS+Wkm+i,r.

[0197] The waiting time summarises the complete busy period until packet number km+i, starts being serviced and reduces it with the time interval Ii−l,r: Wkm+i,r=NstagTS·qkm+i,r−Ii−l,r: Ii−l,r starts at the beginning of the busy period until the rth packet network arrival instant: Ii-1,r=Yi−l+(r−1) NFTF, where Yi−ldenotes an Erlang distributed time interval.

[0198] The total number of network packet arrivals from the begin of the busy period m until service begin of the observed packet is qkm+i,r.

[0199] The total inter-arrival time of the rth packet of the lth call is Ii−l,r=Yi−l+(r−1) NFTF

[0200] The relative total arrival time of the rth packet of the lth call is ΔIi−l,r=(r−1) NFTF

[0201] The number of l=1, . . . , j and j=1, . . . competing packet arrivals between lth connection arrival instant and network arrival instant of packet r from connection l is 7$x_{k_m+i+l,r}^{\min \left(p_{j,r}\right)}=\min \left\{\stackrel{\_}{x}\text{;}r-\lfloor \frac{Y_{l-1}}{N_F{T}_F}\rfloor \right\}.$

[0202] The number of additional packet arrivals of previous connections between lth connection arrival instant and network arrival instant of packet r from connection j (j=2, . . . , i) is 8$x_{k_m+j-1,r}^{\min \left(p_{j,r}\right)}=\min \left\{\stackrel{\_}{x}-1\text{;}\lfloor \frac{Y_{i-j+1}}{N_F{T}_F}\rfloor +r-1\right\}.$

[0203] The amount of additional packets from calls arriving after the observed connection i until network arrival instant of packet number r is 9$x_{k_m+i,r}^{\min \left(p_r\right)}=\sum _{j=1}^{\infty }{p}_{j,r}\sum _{l=1}^j{x}_{k_m+i+l,r}^{\min \left(p_{j,r}\right)}.$

[0204] Number of additional packet arrivals of previous connections between lth connection arrival instant and network arrival instant of packet r from connection i is 10$x_{k_m+i,r}^{\min }=\sum _{j=2}^j{x}_{k_m+j-1,r}^{\min }.$

[0205] The Erlang distributed time interval Yi−l(λ)=Σk=li−lAk(λ) is calculated by composition technique out of i−1 negative-exponentially distributed successive inter-arrival time intervals by generating U1, U2, . . . , Ui−1 (mutually) independent and uniformly distributed between 0 and 1, 11$Y_{i-1}\left(\lambda \right)=-\frac{1}{\lambda }\ln \left(U_1·U_2\dots U_{i-1}\right).$

[0206] pj,r is the probability of j Poisson arrivals during packet producing time interval (r−1) NFTF of connection l, hence 12$p_{j,r}=\frac{{\left(\lambda \left(r-1\right)N_F{T}_F\right)}^j}{j!}{e}^{-\lambda \left(r-1\right)N_F{T}_F}.$

[0207] The mean delay of an arbitrary packet is

[0208] {overscore (d)}=NstagTS+{overscore (w)}=NstagTS+NstagTS·({overscore (x)}−1+{overscore (q)})−{overscore (I)}.

[0209] The mean delay of the rth packet is {overscore (dr)}=NstagTS+{overscore (wr)}=NstagTS+NstagTS{overscore (qr)}−{overscore (Ir)}.

[0210] The mean absolute relative delay of an arbitrary packet {overscore (Δd)}={overscore (Δw)}=|NstagTS·{overscore (x)}·{overscore (Δq)}−{overscore (ΔI)}|.

[0211] The mean delay of an arbitrary packet is the average over all NIP packet delays observed during m=1, . . . , M busy periods: 13$\begin{array}{c}\stackrel{\_}{d}=\frac{1}{\stackrel{\_}{x}·N_{\mathrm{IP}}}\sum _{r=1}^{\stackrel{\_}{x}}\sum _{m=1}^M\sum _{i=1}^{n_m}{d}_{k_m+i,r}=\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}{\stackrel{\_}{d}}_r=N_{\mathrm{Stag}}{T}_S+\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}{\stackrel{\_}{w}}_r\\ =N_{\mathrm{stag}}{T}_s+N_{\mathrm{stag}}{T}_s·\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}{\stackrel{\_}{q}}_r-\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}{\stackrel{\_}{I}}_r\\ =N_{\mathrm{stag}}{T}_S+\stackrel{\_}{w}=N_{\mathrm{stag}}{T}_s+N_{\mathrm{stag}}{T}_s·\left(\stackrel{\_}{x}-1+\stackrel{\_}{q}\right)-\stackrel{\_}{I}.\end{array}$

[0212] The mean delay of the rth packet is the average over all 14$N_{\mathrm{IP}}=\sum _{m=1}^M\left(n_m\right)r^{\mathrm{th}}$

[0213] packet delays observed during m=1, . . . , M busy periods 15$\begin{array}{c}{\stackrel{\_}{d}}_r=\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}{d}_{k_m+i,r}=N_{\mathrm{stag}}{T}_s+\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}{W}_{k_m+i,r}\\ =N_{\mathrm{stag}}{T}_s+\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}{N}_{\mathrm{stag}}{T}_s·q_{k_m+i,r}-\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}\\ =N_{\mathrm{stag}}{T}_s+{\stackrel{\_}{w}}_r=N_{\mathrm{stag}}{T}_s+N_{\mathrm{stag}}{T}_s·{\stackrel{\_}{q}}_r-{\stackrel{\_}{I}}_r\end{array}$

[0214] The mean absolute relative delay of an arbitrary packet is the average over all {overscore (x)}·NIP relative absolute packet delays observed during m=1, . . . , M busy periods is given by 16$\begin{array}{c}\stackrel{\_}{\Delta d}=\frac{1}{\stackrel{\_}{x}·N_{\mathrm{IP}}}\sum _{r=1}^{\stackrel{\_}{x}}\sum _{m=1}^M\sum _{i=1}^{n_m}\&LeftBracketingBar;d_{k_m+i,r}-d_{k_m+i,1}\&RightBracketingBar;=\\ =\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}\stackrel{\_}{\Delta d_r}=\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}\stackrel{\_}{\Delta w_r}=\&LeftBracketingBar;N_{\mathrm{stag}}{T}_s·\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}\stackrel{\_}{\Delta q_r}-\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}\stackrel{\_}{\Delta I_r}\&RightBracketingBar;\\ =\stackrel{\_}{\Delta w}=\&LeftBracketingBar;N_{\mathrm{stag}}{T}_s·\stackrel{\_}{x}·\stackrel{\_}{\Delta q}-\stackrel{\_}{\Delta I}\&RightBracketingBar;\end{array}$

[0215] Average number of cumulative network packet arrivals at network arrival instant of packet number r is {overscore (qr)}=r−1+1/2(NIP−1)+{overscore (xrmin(pr))}

[0216] and for arbitrary network packet arrival instants {overscore (q)}=1/2({overscore (x)}−1)+1/2(NIP−1)+{overscore (xmin)}+{overscore (xmin(p))}.

[0217] Average relative number of cumulative network packet arrivals at network arrival instant of packet number r is {overscore (Δqr)}=r−1+{overscore (xrmin(pr))} and at arbitrary packet arrival instants 17$\stackrel{\_}{\Delta q}=\frac{\stackrel{\_}{x}}{\stackrel{\_}{x}-1}\&LeftBracketingBar;\stackrel{\_}{q}-{\stackrel{\_}{q}}_1\&RightBracketingBar;=\&LeftBracketingBar;\frac{\stackrel{\_}{x}}{2}-\frac{\stackrel{\_}{x}}{\stackrel{\_}{x}-1}\stackrel{\_}{x^{\min \left(p\right)}}\&RightBracketingBar;.$

[0218] Average minimum amount of additional packets from previous connections at network arrival instant of packet number r 18$\stackrel{\_}{x_r^{\min }}=1/N_{\mathrm{IP}}\sum _{m=1}^M\sum _{i=1}^{n_m}{x}_{k_m+i,r}^{\min }$

[0219] and of an arbitrary packet 19$\stackrel{\_}{x^{\min }}=1/\stackrel{\_}{x}\sum _{r=1}^{\stackrel{\_}{x}}\stackrel{\_}{x_r^{\min }}.$

[0220] Average amount of additional packets from calls arriving after the observed connection i until network arrival instant of packet number r is 20$\stackrel{\_}{x_r^{\min \left(p_r\right)}}=1/N_{\mathrm{IP}}\sum _{m=1}^M\sum _{i=1}^{n_m}{x}_{k_m+i,r}^{\min \left(p_r\right)}$

[0221] and an arbitrary network packet arrival instant 21$\stackrel{\_}{x^{\min \left(p\right)}}=1/\stackrel{\_}{x}\sum _{r=1}^{\stackrel{\_}{x}}{x}_r^{\stackrel{\_}{\min \left(p_r\right)}}.$

[0222] Mean total inter-arrival time of an arbitrary packet is 22$\stackrel{\_}{I}\left(\lambda ,\stackrel{\_}{x},N_F{T}_F\right)=\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}\stackrel{\_}{I_{i-1}}=\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}\stackrel{\_}{I_r}=\stackrel{\_}{Y}\left(\lambda \right)+\frac{\left(\stackrel{\_}{x}-1\right)}{2}·N_F{T}_F$

[0223] and for the lth call: 23$\stackrel{\_}{I_{i-1}}\left(\lambda ,\stackrel{\_}{x},N_F^{\prime }{T}_F\right)=\frac{1}{\stackrel{\_}{x}}\sum _{r=1}^{\stackrel{\_}{x}}{I}_{i-1,r}=Y_{i-1}\left(\lambda \right)+N_F{T}_F\frac{\left(\stackrel{\_}{x}-1\right)}{2}$

[0224] and for the rth packet: 24$\stackrel{\_}{I_r}\left(\lambda ,N_F{T}_F\right)=\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}{I}_{i-1,r}=\stackrel{\_}{Y}\left(\lambda \right)+N_F{T}_F\frac{\left(r-1\right)}{2}$

[0225] Mean value of the relative absolute total inter-arrival time of an arbitrary packet: 25$\stackrel{\_}{\Delta I}\left(\lambda ,\stackrel{\_}{x},N_F{T}_F\right)=\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}\stackrel{\_}{\Delta I_{i-1}}=\frac{1}{\stackrel{\_}{x}-1}\sum _{r=2}^{\stackrel{\_}{x}}\stackrel{\_}{\Delta I_r}=N_F{T}_F\frac{\stackrel{\_}{x}}{2}$$\mathrm{and}\mathrm{the}{i}^{\mathrm{th}}\mathrm{call}\text{:}$$\stackrel{\_}{\Delta I_{i-1}}\left(\lambda ,\stackrel{\_}{x},N_F{T}_F\right)=\frac{1}{\stackrel{\_}{x}-1}\sum _{r=2}^{\stackrel{\_}{x}}\Delta I_{i-1,r}=N_F{T}_F\frac{\stackrel{\_}{x}}{2}=\stackrel{\_}{\Delta I}$$\mathrm{and}\mathrm{the}{i}^{\mathrm{th}}\mathrm{packet}\text{:}$$\stackrel{\_}{\Delta I_r}\left(\lambda ,N_F{T}_F\right)=\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}\Delta I_{i-1,r}=N_F{T}_F\left(r-1\right).$

[0226] The mean value of NIP Erlang-(i−l) distributed time intervals is given by 26$\stackrel{\_}{Y}\left(\lambda \right)=\frac{1}{N_{\mathrm{IP}}}\sum _{m=1}^M\sum _{i=1}^{n_m}{Y}_{i-1}.$

[0227] The Hypo-exponential Process is here used to construct a substitute probability distribution function and consists of a discrete time process D with random variable Tl and mean tl={overscore (d)}·(1−c) linked with a negative exponential process M with random variable T2 and mean t2={overscore (d)}·c Ftt O for O<t<tl FD(t;tl 2)=if e-(t) lt2 for t>t,

[0228] The probability distribution function of the Hyper-exponential Process is used to construct a substitute probability distribution function and is given by

[0229] FD(t, p, t1,t2)=1−p·e−(t/t1)−(1−p)·e−(t/t2)

[0230] with the mean values 27$t_{1,2}=\stackrel{\_}{d}·{\left(1\pm \sqrt{\frac{c^2-1}{c^2+1}}\right)}^{-1}\mathrm{and}\mathrm{probability}p=\frac{\stackrel{\_}{d}}{2·t_1}.$

Claims

1. Method for reconstructing non-continuous packetized data of a continuous data stream like streamed media, voice, audio, or video from a data connection into a continuous data stream at the receiving point of a packet-based network, comprising the steps of providing of at least one estimation method based on at least one characteristic value concerning data connections of the kind intended for, gathering measurements of at least one value characterizing the data connection, evaluating a de-jittering delay for the data connection by predicted parameters taking into account the at least one provided value and the at least one gathered value, delaying and converting the data packets following the evaluated de-jittering delay.

2. Output unit for reconstructing non-continuous packetized data of a continuous data stream like streamed media, voice, audio, or video into a continuous data stream, comprising means for providing of at least one estimation method based on at least one characteristic value concerning data connections of the kind intended for, gathering measurements of at least one value characterizing the data connection, evaluating a de-jittering delay for the data connection by predicted parameters taking into account the at least one provided value and the at least one gathered value, delaying and converting the data packets following the evaluated de-jittering delay.

3. Terminal with output unit for reconstructing non-continuous packetized data of a continuous data stream like streamed media, voice, audio, or video) into a continuous packet data stream, comprising means for providing of at least one estimation method based on at least one characteristic value concerning data connections of the kind intended for, gathering measurements of at least one value characterizing the data connection, evaluating a de-jittering delay for the data connection by predicted parameters taking into account the at least one provided value and the at least one gathered value, delaying and converting the data packets following the evaluated de-jittering delay.

4. A computer program product including software code portions for performing steps of claim 1 for a terminal.

5. A computer program product including software code portions for performing steps of claim 1 for an output unit.

6. A method according to in claim 1, wherein the packet inter-arrival times are measured.

7. A method according to in claim 1, wherein the variation of the packet inter-arrival times is calculated.

8. A method according to in claim 1, wherein a network description and routing information for predicting said parameters are used.

9. A method according to in claim 1, wherein an initialization based of measurements of at least one value characterizing the data connection gathered by the first transmitted packets is performed.

10. A method according to in claim 1, wherein an initialization based of measurements of at least one value characterizing the data connection gathered by previous transmissions is performed.

11. A method according to in claim 1, wherein an initialization based of measurements of at least one value characterizing the data connection gathered by previous connections is performed.