FANS Simulation of Optical Burst Switching for NSFNET

M. El Houmaidi, O. Kachirski, R. Guha School of Computer Science University of Central Florida, Orlando, FL 32816

Abstract :- In this paper we propose a simulation framework designed for simulating advanced network technologies. We provide a detailed architecture of the framework and describe all the necessary steps to develop simulations for any particular networking application under consideration. We demonstrate the use of this framework by simulating an optical burst switched network to aid in our quality of service studies. With the help of simulation results we show that our QoS-enhanced signaling protocol for optical networks achieves better performance and lower burst loss than existing protocols.

1. Introduction placement algorithm using limited range conversion. The complexity of the protocol and the numerous It has been established that Internet traffic is self- constraints achieving the optimal performance, similar, bursty and has a heavy tail statistical nature dictated the need for a robust, extensible simulation [4]. This traffic pattern-based observation leads to the framework to further study possible improvements to need for a signaling protocol not only suitable for the Quality of Service (QoS) in BSON. long duration flows but also improving the network throughput and the provisioning speed at the same FANS framework (Framework for Advanced time. Optical burst switching with its signaling Network Simulation) was developed by our research protocol suite is the switching paradigm of choice group at the University of Central Florida as a combining the best of Optical Flow Switching and research tool to provide time-saving flexible test-bed Optical Packet Switching for high bit-rate transfer, for a wide range of network technologies. The providing real-time support for multi-media internet framework is cross-platform, GUI-driven and can be traffic, and fast switch configuration for short used as a generic advanced network simulator for a duration sessions. variety of purposes, such as routing in ad hoc networks, QoS schemes for burst switched optical The latest signaling protocol, JET (Just Enough networks, mobility models of totally mobile wireless Time) [1] for burst switched optical networks networks, and Bluetooth Pico-nets. We currently use (BSON) controls the end-to-end delay by adjusting an implementation of FANS architecture as a test-bed the offset time at the source edge node based on the for our burst switched optical network algorithms processing time of the control packet along the path testing. In the following sections we first describe the taken by the burst. The burst is buffered at the source architecture of FANS framework, then give an edge node for the duration of offset time. JET does example of simulating QoS combining Fiber delay not support any QoS scheme taking into lines and Optical wavelength converters for burst consideration the properties of the WDM switched optical networks using FANS. (Wavelength Division Multiplexing) layer. Furthermore, our scheme combines the contention resolution with a dynamic class-based allocation of 2. FANS architecture resources such as FDLs (Fiber Delay Lines) and We have developed a customized network OWCs (Optical Wavelength Converters) [2]. simulator that is general enough to target advanced networks of diverse configurations, yet is easily This new switching paradigm is continuously configurable for specific network applications. Our evolving and all its components (signaling, goals were to create a clean, easy-to-understand and burstification, routing and QoS) are being studied and extensible design using object-oriented approach, enhanced by the research community. In [2], we have ensure general cross-platform portability, and introduced an optical wavelength converter intuitive GUI-driven simulation interface. Performance and scalability were the two most slow execution times. The FANS framework avoids important design issues. The architecture of FANS these problems, by implementing only the key (simulator and interface) is based on a building-block functionality of a network – supporting simulation approach. Researchers implement algorithms or clock, network communication and node behavior. prototypes from modules that receive inputs in the These functions are implemented on an abstract level, form of events, process them, and generate outputs allowing the user to include specific communication (events, log entries, GUI updates). The entire protocols, routing algorithms and other required network is built from objects – nodes, links, event simulation parameters, as necessary. The intuitive generators, and communication channels. graphical user interface reflects any changes in the Connections between nodes are maintained by the simulation execution as they happen, in real-time, connection handler object. Most common types of allowing the user to adjust the simulation parameters network architectures are supported and can be at run-time. Another advantage of our framework is configured – such as, for example, ad hoc and the ability to use almost any data file as a source of infrastructure wireless networks [5]. Object-oriented network communications. Data file pre-processor approach is central to the generality and flexibility of converts binary network traffic data into the format the system and allows users of our framework to accepted by FANS simulation system, converting reuse, share and catalog simulation components by network addresses into plain addressing scheme used modifying or replacing appropriate classes. with our simulation framework. As an example, intrusion data from the Lincoln Laboratories IDS Majority of existing advanced network tests was used to test our wireless IDS system [6]. simulators aim at supporting every possible aspect of network communications, such as, for example, FANS architecture consists of four key modules,

User Stats Logging Module Stats simulation Collection control

G U I

Data Parser CSV File for object Analysis representation

Network Traffic Simulation Engine Traffic Processor Module

Messaging

Simulation Objects BIN/XML Parser Traffic Data File

Simulator Core Module

Figure 1. FANS Architecture. every layer of many communication protocols used. each comprised of a number of components - GUI This adds tremendous overhead to the simulation (graphical user interface), Simulator Core Module, system, often resulting in scalability problems and Network Traffic Module, Data Collection Module. Figure 1 shows the FANS architecture. Some of the transmitting data (frames/bursts), and has routing modules carry optional functionality and can be information as its arguments. Send method included into the simulation as necessary. The determines the proper route, and sends the traffic functionality of each module is described further on along this route by invoking Receive method on each in this section. node recursively. Receive method first checks the traffic destination, then depending on the routing - Graphical User Interface Module algorithm used, forwards the traffic to the destination The graphical user interface (GUI) module shows or performs other traffic manipulation (such as burst the simulation execution in real-time. The GUI class delay). This allows the user to simulate various is tied up to the simulation engine clock, and displays network architectures – from wireless to optical. the required information every clock cycle. GUI class Routing protocol is another simulation object. The shows the simulation area with simulation objects. routing protocol included with FANS framework is a Certain simulation parameters can be adjusted via simple table-driven protocol, which is implemented GUI module, such as traffic load and simulation as a separate routing class. To modify the routing clock delay. GUI module also provides controls for protocol, user must include all necessary parameters simulation execution – user can pause simulation, in a network node object, then add their own routing step through the simulation one clock cycle at a time, protocol class. add and remove simulation objects. GUI module also records simulation statistics in a log window. Data - Network Traffic Module logging module processes this data at the end of Network traffic for FANS framework is simulation and saves the results in CSV format, generated by the Network Traffic Processor object. which can be later processed by a mathematical The implementation of the object is common to all analysis software package. the simulations and includes reading a pre-processed data file in XML format, and routing network traffic - Simulator Core Module to the appropriate network node. FANS uses a flat The simulator core module is composed of the addressing scheme to reduce communication simulation engine and the simulation objects. overhead, and therefore all network addresses are Simulation Engine: The heart of the FANS converted to compatible notation by the BIN/XML framework is a Simulation Engine. The simulation is Parser module before running the simulation. Traffic tied up to a graphical user interface, allowing users to processor object generates network traffic at certain monitor simulation progress in real-time and modify timestamps, and makes use of simulation engine’s simulation parameters on the fly. The simulation internal clock to time distribution of events. This engine runs an internal clock, the speed cycle of allows us to speed up simulation execution, limiting which can be controlled at run-time. the simulation speed only by the hardware Programmatically implemented as a high-priority specifications and the maximum incoming event rate. thread, the simulation engine runs in a loop continuously, driving execution of all other - Data Logging Module components of the simulation. Any events happening Data logger class saves the simulation results for at a certain time must be processed by simulation future analysis. During simulation execution, results objects at once, within a single simulation cycle. The are stored in memory and displayed in a readable simulation engine is common to all simulation form via GUI data display window for performance models, and cannot be modified by the user. reasons (frequent disk I/O operations reflects Simulation Objects: Simulation objects execute negatively on simulation performance). The independently and are time-synchronized via the representation of results can be tailored to particulate simulation engine clock. Our FANS framework has a simulation requirements and is defined in the GUI few pre-defined simulation objects. One object type class. is a network node object. Nodes are members of any Data parser: At the end of the simulation run, network simulation, and can be either stationary or these results are first pre-processed by a data parser mobile. Each network node object also includes two to format data suitable for import into the methods used for inter-object communication – Send mathematical analysis software. and Receive. Send method is invoked when a node is CSV file generator: The pre-processed output is models seven different events: an arrival of a control saved as a CSV file (a widely-used comma-separated packet for path establishment in the network, a data file format) by the CSV file generator. departure of a reserve control packet from a particular node to the next hop node, the event of Simulation process blocked burst due to lack of resources, the event of arrival of the burst at the destination node, the release The simulation procedure is as follows. First, a of resources by a control packet, the event of burst simulation diagram is devised by the user, which ready after FDL delay, and the event of ending the includes the list of all objects taking part in the simulation after C bursts are processed (simulation simulation and interaction mechanisms between the stop condition). C is 106 times the considered load of objects. Statistics output formats, as well as traffic in the network. The load, in Erlangs, is the simulation stop conditions should also be specified at product of the arrival rate and the average burst this point. Next, the implementation is written for all length. The simulation takes into consideration the of the above objects, following the guidelines control packet processing time, P, at each node. P is included with the simulation framework. Data parsers assumed to be 10 µs. The propagation delay is should also be adapted to reflect the new data estimated based on the kilometric distance between structures. The objects are then placed in the nodes and the speed of light. Three alternate hop- simulator core directory and compiled. The GUI is based shortest paths between node-pairs are then started, which then instantiates all simulation computed in advance. If the burst is dropped when objects and displays them on-screen (for those routed along the first path, the second, then third objects that have visual behavior defined). Objects alternate paths are tried. then can be manipulated via graphical interface during simulation runtime. Events are passed to 15 appropriate objects via messages and processed. All 12 collected statistics are displayed in a log window, 11 which is then saved into a file. The results of the 3 9 simulation can be pre-processed for further analysis 14 or graph plotting by the Data Parser module. The 8 supplied data parser converts log file into a CSV 1 7 format, which then can be imported into a variety of 4 6 13 statistical programs. 0

2 5 10 3. FANS for burst switched optical

networks design Figure 2. NSFNET Topology.

Our first use of FANS architecture was to develop a test-bed for a wireless intrusion detection We consider the NSFNET topology (figure 2) which system (IDS) simulation prototype [5, 6]. In this is used extensively in optical networks simulations. paper, FANS architecture is applied to burst switched We show the benefits of QEJET supporting 4 classes optical networks (BSON). This section covers the of traffic. Each fiber link has W wavelengths. The architecture of our simulation, step-by-step design burst length is exponentially distributed with an process, implementation of several BSON-related average of L μs and the normalized load intensity is algorithms, and simulation runs. ρ.

Our FANS simulation is based on a discrete- event model with a time-advance mechanism to the most imminent event to be processed. Each event We assume that FDLs and OWCs are available at processing affects the state of the system and the time selected core nodes of the OBS network. OWCs are of occurrence of future events. Our simulation used for high priority bursts where end-to-end delay needs to be kept to a minimum. FDLs are used to Burst loss probability vs. load under JET and QEJET delay a burst at a core node to resolve contention for supporting 4 classes of traffic (C=4) using W=4 higher priority bursts. High priority bursts take Load intensity advantage of wavelength conversion to incur minimal 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 end-to-end delay and lower priority bust will be 1.E+00 delayed using a variable-delay FDL bank (each output fiber has a dedicated FDL). As in [3], we adopt the variable-delay FDL design but our design is 1.E-01 different than [3] in term of usage of wavelength QEJET Class 0 converters to allow the maximum possible of bursts

y 1.E-02 t i l to be delayed at the same time. Each fiber delay i QEJET Class 1 b a b o element (represented by a circle) can provide a delay r p

s

s Classless JET of b time units. A delay line can provide a variable o l

1.E-03 t

0 1 s r delay ranging from zero to MAXD = (2 + 2 + … u b +2(max_d)) x b units of time. QEJIT Class 2(W=4) 1.E-04 Our QoS enhanced JET scheme is based on the QEJIT Class 3(W=4) Horizon scheduling [1]. The algorithm has an advantage in hardware implementation. Each class of 1.E-05 traffic has a priority that drives the allocation of FDLs and OWCs. Higher priority bursts require Figure 3: QEJET: burst loss versus load (W=4) minimal end-to-end delay and will use OWCs as much as possible (pre-emption is possible of lower priority bursts). Low priority bursts will incur delays Burst loss probability vs. load under JET and QEJET supporting 4 classes of traffic (C=4) using W=64 in FDLs. Load intensity We assume that certain switches are equipped 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 with FDLs and OWCs, max_d is 8 and consequently 1.E+00 MAXD, representing the maximum delay, is 255xb μs 1.E-01 and the transmission rate is 10 Gbps, the cross- connect setup time is 10xb μs. The considered 1.E-02 performance metric is the burst loss probability as a QEJET Class 0 y function of the load intensity ρ and the number of t i l i 1.E-03 QEJET Class 1 b a

wavelengths W. Figure 4 and 5 show the relationship b o r p

between the burst loss probability and the load s s 1.E-04 Classless JET o l

t intensity ρ ranging from 0.2 to 0.9. We fixed burst s r u b length to L = 100xb and b = 1 μs. 1.E-05 QEJET Class 2

We measured the experienced burst loss as a 1.E-06 QEJET Class 3 probability for classless JET (char in blue); and also when QEJET is supporting 4 classes of traffic (C=4), 1.E-07 the burst loss for each class (i=0..3) is measured for W=4 in figure 3 and W=64 in figure 4. All classes of traffic have equal mean burst length, i.e. Figure 4: QEJET: burst loss versus load (W=64)

L3=L2=L1=L0=L; and generate an equal amount of traffic. Our results show that QEJET achieves a class Indeed, with 4 wavelengths per fiber, we observe, on differentiation in terms of incurred burst loss: class 3 one hand, that class 3 under QEJET has lower burst experienced a reduction of burst loss as compared to loss probability than the classless JET for all values the classless JET: 50 and 600 orders of magnitude of the load intensity. On the other hand, class 2 when W=4 and 64 respectively. achieves lower burst loss than the classless JET only when the load intensity is higher than 0.4. We also note that class 0 and 1 experience higher burst loss [1] C. Qiao and M. Yoo, “Optical burst switching than the classless JET scheme for all considered (OBS) – a new paradigm for optical internet”, loads. When we use 64 wavelengths per fiber, we Journal of high speed networks, vol. 8:1, pp. 69 notice that the QoS performance can be significantly -84, 1999. improved by increasing the number of wavelengths: Only class 0 incurred higher burst loss than the [2] M. El Houmaidi, M. Bassiouni and G. Li, classless JET, Class 2 and 3 for all loads and class 1 “Limited range wavelength conversion in WDM when the load is higher than 0.6. optical networks”, OISE 03, Optics in the Southeast conference, (CREOL, OSA, SPIE) November 2003. Summary We have introduced the FANS architecture for [3] C. Hsu, T. Liu and N. Huang, “Performance of advanced network simulations. It has been developed adaptive routing in wavelength-routed networks to aid our research on Intrusion Detection Systems with consideration of delay-sensitivity”, for Wireless Networks, and later used as a research Twenty-First Annual Joint Conference of the tool that incorporates a flexible test-bed targeting IEEE Computer and Communications Societies, simulations of a wide variety of advanced network INFOCOM proceedings, vol. 1, pp. 66 -73, June communications. FANS framework has been used to 2002. validate our QoS scheme for BSON, as described in this paper. The simulation results show that our QoS [4] V. W. S. Chan, “Some thoughts on optical enhanced signaling protocol achieves a class networks”, MIT, DOD-N presentation, differentiation in terms of incurred burst loss. http://www.darpa.mil/mto/solicitations/BAA03- Furthermore, the QoS performance can be 19/S/presentations/mit_chan.pdf (2003) significantly improved by increasing the number of available wavelengths. [5] Kachirski, R. Guha, “Effective Intrusion Detection Using Multiple Sensors in Wireless The FANS framework is cross-platform, Ad Hoc Networks”, Proceedings of 36th HICSS scalable, and flexible to adjust to particular Conference, 2003, pp. 57-64. requirements, such as simulating BSON. It is undergoing significant development, and we are [6] R. Guha, O. Kachirski, D. G. Schwartz, S. working with the research community on adding all Stoecklin, E. Yilmaz, “Case-Based Agents for the necessary features for FANS to establish itself for Packet-Level Intrusion Detection in Ad Hoc a wide-spread use in computer network simulations. Networks”, Seventeenth International Symposium On Computer and Information Sciences, Orlando, FL, October 28-30, 2002 Acknowledgements This work has been supported by NSF under Grant No. 0114418 and Grant No. 9980316, and the US Army Research Office Grant No. DAAD19-01-1- 0502. The views and conclusions herein are those of the authors and do not represent the official policies of the funding agencies.

References