Chapter VI

CHAPTER VI

C ONCLUSION AND F ^UTURE D ^IRECTIONS

This thesis has been accomplished in the context of the OSISIM project. The main contribution is the emphasis made on the reusability concept, on one side, for designing a simulation environment, and on the other side, for defining two different levels of granularity for reusable network component libraries. In retrospect, one can clearly see the positive potential influence that reusability is having on the development of high quality software and models.

The design of our simulation environment, called AMS, was based on existing pieces of software, which proved their usefulness in their respective fields. In order to carry out this integration efficiently, a modular structure of the atelier was proposed. The structure has been divided into four phases. Each phase is responsible of a part of the performance evaluation cycle. The main novelty of this structure is the usage of a dedicated language as a means to define a clear border between the editing and simulation phases and to allow the portability of the atelier upon different platforms. During the specification of the Architecture Description Language (ADL), we took into account the fact that this language has to be easy to learn and to use, has to have the capacity to hide simulation details from end-users and AMS-managers, and has to have the main characteristics of a modern language such as object-orientation and modularity.

A prototype of the atelier has been developed on a SUN machine running the

SunOs operating system. It is developed in C language. At present time, about 15

000 lines have been coded. It was presented in several national and international public events.

The kernel of the AMS is its library of Detailed Basic Models (DBMs). Each DBM was designed in order to comply with the most important criterion which is reusability. Indeed, each DBM can be used in several network architectures and can be a component of generic and composite models. Before the effective usage of a DBM, it is verified and validated in order to increase the model credibility.

Chapter III was dedicated to the description of a basic model, its internal structure, and how it can be connected to other basic models. A DBM has been structured in three blocks: the Behavior Engine Block, the Interfaces Block and the Measurements Block. The concepts of facet, superfacet, and mutual exclusion were defined in order to apply coherently and efficiently the composition mechanism. Indeed, using the composition mechanism, a modeler can define generic and composite models.

Several DBMs are already developed such as MAC FDDI [19, 20] and MAC DQDB [19, 23, 24]. In [22], details are given about MAC Ethernet, TCP, FTP, and a Bridge as a generic model.

In order to assist the modeler in his work, new tools have to be added to the AMS such as a DBM editor to help the modeler to write models complying with the structure of a DBM and respecting the main pre-defined guidelines, or to apply the composition mechanism between existing DBMs.

Even with this tool, a modeler has to spend a long time for modeling a network component because there is no methodology to help him to address this problem.

The most important contribution of this research is the definition of a methodology for modeling protocol entities as DBMs. We then tried to partly bridge the gap between specification and modeling. Chapter IV presents this methodology which is based on the concept of function. Simple functions are modeled as reusable modules and stored into a library.

The Function Based Methodology was designed to help the modeler to build

efficiently and rapidly new protocols designed for the new generation of

networks where several services can be provided. These new protocols can be

dynamically tailored to the user’s requirements. The methodology is composed

of several steps, the most important are : i) modeling simple functions as

modules with known sets of inputs and outputs, ii) partitioning the library of

functions into several parts depending on the number of event classes handled by the modeled protocol entities, iii) defining the precedence graphs associated with these parts, iv) defining the -Graph, and v) executing PPDP to determine peer- protocol entities patterns providing selected services.

We have shown, in chapter V, the usefulness of the Function Based Methodology to model a recent high-speed transfer protocol, XTP. The building of the model was almost straightforward when all the steps defined by the methodology are followed. The model is reconfigurable and several services can be provided such as stream and isochronous stream services. On the other hand, a complete implementation of the OSI TP model, structured as a DBM, was done following this methodology [21].

Several tools have to be developed to help the modelers to apply efficiently and coherently the Function Based Methodology, such as coding the function modules, defining the precedence graphs and the -Graphs, and generating the protocol entities patterns.

F U T U R E D I R E C T I O N S

The work achieved during this dissertation can be enhanced in many directions.

Three major research directions have emerged during our reflection and our several readings.

Reusability and Hybrid Simulation Modeling

The main disadvantage of simulation is the slowness of the execution. One solution to this problem is to apply hybrid simulation modeling (HSM) techniques [98, 99]. The goal of HSM is to build models which are, on one side, more representative than pure analytical models and, on the other side, that lead to a substantial reduction of execution time with respect to pure simulation models. The HSM techniques were already used for particular situations but reusability was not taken into account.

In order to define more formally HSM, we refer to the classification introduced

in [100], that distinguishes four classes of HSM by considering four different

interaction ways between simulation and analytical models. Specifically, two of

these classes (I and II) include the combination over time of simulation and

analytical solution either in parallel or through a joint solution procedure. The other two classes (III and IV) consider either a pure analytical or simulation model of the total system and use, respectively, a simulation or an analytical solution to represent a portion of the system.

In our case and for reusability purpose, the fourth class of HSM has to be studied more thoroughly. Actually, the total system can only be resolved by simulation with one or several parts modeled analytically.

Another solution to reduce the execution time of a simulation is to distribute the simulation over parallel processors or several machines. As DBMs are autonomous processes, parallel or distributed simulation [101] may help to overcome this problem.

Reduced Basic Model

Another issue which remains for further research study is how to reuse network components where each component is modeled at different levels of detail.

During the development of this dissertation, we have considered only detailed models. On the other words, each basic model is detailed so as to reflect its exact behavior. The reasons why we have chosen this approach are that we want to have a model behavior close to a real behavior, and to be able to derive accurate measurements from the simulation runs.

However, this approach has some drawbacks because of the large size of the code resulting from the description of the DBM. The main drawbacks are : i) the simulation is very time consuming, ii) the code describing a system is, of course, very long, because the system is composed of several DBMs, each of them having a long code, iii) during the simulation run, a large-sized memory is needed, and v) the end-user will be submerged with details, so that his editing work will be difficult.

In order to soften the effects of these drawbacks, we suggest reducing the DBM

code, by eliminating, for example less significant details. Then, the code obtained

will be called RBM for Reduced Basic Model. To do this, we are faced with new

problems, such as : how can less significant details be chosen ? and during the

construction of the system code, the question that arises is : for which component

is the DBM required ? and for which one is the RBM sufficient ?; the choice may

be made even more difficult, because each DBM can have several RBMs associated with it.

Traffic Modeling

Modeling properly the data stream traffic is a crucial part of any performance evaluation study of a communication system. Since it is possible to reach misleading conclusions if unrefined traffic models are used, generally, data traffic is modeled as Poisson processes. However, recent traffic studies have shown that packet inter-arrival times are not exponentially distributed [102, 103, 104].

To address this problem, the analysis of real traffic has to be carried out in order to build traffic models. A certain number of works have already been accomplished either for collecting information from real networks or for analyzing and building models. However, there is more work to be done in understanding traffic patterns, and a better understanding of these should impact the design of future networks.

Radouane MRABET

March 28, 1995.