The generic framework targets multimedia application modelling as well as IP traffic modelling. Two issues were considered when designing the tool: The packet transport protocols and the application behaviour. From a traffic modelling point of view we have two types of packet generation profiles: deterministic with UDP packets and closed-loop controlled with TCP packets. On the other hand, multimedia applications on packet switched networks share a common behavioural structure composed of active and idle periods. Thus, despite the diversity of multimedia applications and traffic profiles they may produce, they can be modelled as a generalized ON-OFF process. Indeed, audio, video and data applications are composed of a succession of active and idle periods. According to the complexity of modelled applications we may have a simple periodic succession of ON and OFF periods or a multilevel non-periodic structure of ON-OFF processes. In order to take into account all possible behaviours we suggest a hierarchical representation of multimedia applications (and traffic sources in general). Thus, multi- media application behaviour can be recreated with three levels: Session, Activity and Packet.
• Session Level: It models the arrival of clients connecting to the system in order to use a given application.
• Activity Level: It models the density of exchanged information accross the time. This could be different depending on the application structure.
• Packet Level: It is the basic level at which packets are generated according to information exchanged during the application.
Chapter 3. Generic Framework for Traffic Modelling 37
3.4.1
Session Level
A Session level models the arrival process of clients into the system. It is the behaviour of the population of users at the access node (router on a LAN, Base station in radio networks . . . ). Two cases can be considered for client arrival process:
• Constant number of sessions: the number of connected users is fixed and does not change.
• Random number of sessions: sessions are established randomly and sessions’ dura- tions are application dependent.
Of course when the number of sessions is constant there is no arrival process to define, and session durations are considered illimited. The constant number of sessions model is very important to evaluate the performance of a system under constant number N of applications. On the other hand, the random number of sessions model involves two processes: the session arrival process and the session duration process.
The session arrival process can be of any type, but most common models are: • Poisson model: It is the most popular model used to represent the client arrivals.
This model represents an average rate of exponential arrivals λ, without memory, independent of previous arrivals.
• Truncated Poisson model (or Erlang model): Clients arrivals are still considered exponential but only a maximum number of active sessions is allowed. It is generally expressed in terms of blocking probability per type of session. This model is widely used in PSTN networks and still has applications in multiservice networks.
The session duration process depends on the transport protocols. In real time ap- plications, the duration of a session is characterized by a distribution depending on the type of application. While in non real time applications, the communication time is controlled by TCP. In fact, in such case the source cannot determine the temporal se- quence of packet inter-arrivals. This type of sources is better modelled by the quantity of information to transmit. The duration of the session depends on the network response and not on the source itself.
3.4.2
Activity Level
The density of information of an application is specific to each application. To describe this property, the activity level represents the application as a set of parallel flows. Each flow is composed of a succession of active and idle periods. The main behaviours are listed below:
• ON-OFF Behaviour in a conversational process (HTTP, VoIP. . . ) where ON period represents the active period during which packets are transmitted and OFF period represents idle time.
Chapter 3. Generic Framework for Traffic Modelling 38
• ON behaviour during all the session time, such as FTP file transfers. This behaviour may be used with complex processes like correlated event processes, in order to define aggregated traffic models.
Active periods (denoted ON) can represent any kind of activity, using the probability distribution of packet inter-arrivals and sizes. ON periods can also describe more complex behaviours by subdividing the ON period itself to sub-ON periods. The Web model, for example, can be described by simple ON period representing the page downloading time followed by an OFF period representing the reading time. ON and OFF periods are grouped in one entity called Pattern. In this case the pattern is the Web page and the occurrence of the pattern defines the repetition of the Web pages. However, the Web model can detail the activity in the ON period. Hence, the page downloading period can be divided into sub-ON periods to reflect the presence of inline objects (image, applet, video . . . ). See Chapter 5 for more details.
The activity level handles another specific characteristic of multimedia applications which is parallel flows. Actually, different parallel flows may be present during application life time. Those flows concern main application activity and signalling protocols flows. For example, a voice or video application may be transported by RTP and controlled by RTCP.
3.4.3
Packet Level
This is the basic level of the hierarchical model. At this level one defines the way packets of the application will be generated and transported. If the UDP protocol is used then one determines the packet inter-arrival and size distributions. The packet inter-arrivals may follow any kind of distributions, while packet sizes can only follow truncated versions of distributions to respect packet size limits on the networks (e.g Internet, Ethernet . . . ). If TCP is used, packet inter-arrivals are determined by the TCP automata according to the reception rate of ACKs. Figure 3.3 illustrates the three levels.
The generic framework represents the core of a complete tool conceived for the design and the evaluation of multimedia traffic sources in network environment called Traffic Source Modeler (TSM). A more detailed description of the tool is given in Appendix A.
3.4.4
Packet Rate Estimation
Some multimedia applications are very complex. As a consequence, the correspond- ing application models conceived with the generic framework could be very sophisti- cated. Several flows with real time and non real time traffic patterns alternating could be present. For example in new chat sessions, end users exchange text, audio and video data. Hence, a simple evaluation tool for a new application model is to determine its average packet rate. Unfortunately, the packet rate estimation is not possible with TCP flows. As we know TCP has an elastic behaviour that depends on network congestion, thus the estimation of packet rate in this case is problematic. To overcome this difficulty
Chapter 3. Generic Framework for Traffic Modelling 39
Figure 3.3: Three Level Description of Multimedia Applications
we suppose that the TCP source rate is bounded by terminal packet rate on which the multimedia application will be deployed. Indeed, this assumption is justified because TCP sources stabilize after the transient phase on a nominal packet rate normally equal to the terminal packet rate. Of course, packet rate estimation in this case is approxi- mate but still useful to gain insight on possible packet rate generated by the application when deployed on the network. In the following we present the estimation of the average packet rate of a multimedia application conceived with the framework.
Generally, application models have different flows (streams). Each flow has patterns composed of ON and OFF periods. Consider:
• The number of streams is Ns.
• The number of repetitions of pattern m in stream s is Nm,s.
• The duration of pattern m in stream s is Tm,s.
• The packet rate of the period p in pattern m in stream s is λp,m,s.
• The duration of period p in pattern m in stream s is Tp,m,s.
• The period p file size (when it applies) in pattern m in stream s is Qp,m,s.
• The average packet size during period p in pattern m in stream s is Pp,m,s.
• The packet inter-arrival during period p in pattern m in stream s is IAp,m,s.
• The packet rate of stream s is λs.
Chapter 3. Generic Framework for Traffic Modelling 40
We want to estimate the packet rate (denoted λ) of a multimedia application described by the previous parameters.
We have: λ = Ns X s=1 λs (3.1) Also: λs = P m∈Mλm,s∗ Nm,s∗ Tm,s P m∈MNm,s∗ Tm,s (3.2) And: λm,s= P p∈Pλp,m∗ Tp,m P p∈PTp,m (3.3)
Where λp,m,s = IA1 and IA is the average packet inter-arrival in period p of pattern m.
The value of λp,m can not be determined in the case of TCP based traffic sources,
as it is function of network congestion. It is replaced by the nominal value of access terminal packet rate. The duration of a TCP period is given by:
Tp,m,s=
Qp,m,s
Pp,m,s⋆ λp,m,s
(3.4)
The packet rate during period p depends on its type, OFF periods have null value for packet rate:
λp,m,s =
(
λON ON
0 OF F (3.5)
Using the previous formulas we can give an estimation of the average packet rate produced by a multimedia application described by the framework. The estimated packet rate is considered as a characteristic of the multimedia application model and it is saved with its description for later use.