A New Distributed Dynamic Bandwidth Reservation Mechanism to Improve Resource Utilization Simulation and analysis on real network and traffic scenarios Sukrit Dasgupta, J. C. de Oliveira Electrical and Computer Engineering Department Drexel University Philadelphia, PA 19104 Email: {sukrit,jau}@ece.drexel.edu J.-P. Vasseur Cisco Systems Boxborough, MA, USA Email: jvasseur@cisco.com Abstract—In this paper, a novel distributed dynamic MPLS traffic engineering (dynamic TE) mechanism is proposed and its performance under failure scenarios and increased traffic load is analyzed when compared to other traditional approaches (static IP and TE). The new mechanism periodically updates bandwidth reservation and selects path (resizing and rerouting) for each TE LSP according to its computed traffic load, which leads to path reoptimization and better network utilization. Dynamic TEs performance is analyzed via simulation using real network and traffic scenarios in a newly created simulator, which was carefully developed to reproduce as accurately as possible the behavior of such networks. The simulation results show the significant performance improvement of dynamic TE for several metrics of interest and give insight into other scenarios that could benefit from its deployment. I. I NTRODUCTION Network resource allocation has been an area of study for many researchers in different fields of networking. When trying to allocate resources in a shared data center running web applications, researchers have realized the need for dynamic mechanisms, which are able to adapt to time variability of web workloads in order to provide Quality of Service (QoS) guarantees to such applications. Similarly, when dealing with multimedia streaming over the Internet, the research commu- nity has yet again converged to the need of dynamic resource allocation [1], [2], [3]. While applications and scenarios may vary, the search for QoS provisioning brings up the need for resource reservation and to take time-varying network conditions into account. Since network resources are shared, providing guarantees to applications is a complex task and are usually provided by reserving a fraction of network resources for every traffic aggregate. This fraction mostly depends on the expected traffic load and QoS requirements of the application. The workload of web applications, for instance, is known to vary dynam- ically over multiple time scales [4] and can be influenced by unanticipated events such as breaking news or a planned event such as the live 8 performances broadcasted around the globe, etc. Furthermore, it is a well-established fact that network consumption significantly varies during the day, a fact that is exemplified in global networks with multiple time zones. Consequently, static allocation of network resources has its own limitations. While over-provisioning resources based on worst case workload estimates can result in potential underutilization of resources, under-provisioning resources can result in violation of QoS guarantees. These limitations lead to the alternate approach of allocating resources dynamically based on the variations in workload. Several research papers have investigated this proposal [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12]. While many researchers concentrate on forecasting the traffic demand, others propose a centralized controller to provide optimum allocation. In both cases, most works concentrate on the problem of optimally partitioning the traffic into classes and the link capacity into bandwidth partitions to be assigned to each class. Based on on-line mea- surements (or traffic forecasting), the network then combines real-time traffic and network state information in order to optimize the size of the bandwidth partitions or the service rate of the traffic classes to achieve better performance [12]. We investigate the problem of dynamic resizing after traffic classes have been already decided on; resources (reserved bandwidth) have been reserved in the network for each class and a corresponding route has been found in the network for the incoming traffic. In this paper, we propose a novel distributed dynamic Traffic Engineering (TE) mechanism - “Dynamic TE, where the reserved resources are updated dynamically based on the computed load and the route is also updated accordingly to provide the traffic with the shortest available path in that would fit the traffic demand. We select MultiProtocol Label Switching (MPLS) as the underlying technology for imple- mentation due to its efficient traffic engineering granularity and wide deployment by network providers around the globe. In particular, aiming at an accurate study, we consider the following: • Two different topologies, small and large, with various degrees of meshing, both representing realistic networks, • A realistic traffic load, where a large number of small size aggregate traffic competes with a smaller amount of medium and large size aggregate traffic, • Traffic shapes reproducing the expected traffic pattern over time, This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the Proceedings IEEE Infocom. 1-4244-0222-0/06/$20.00 (c)2006 IEEE