Performance modeling and management of high-speed networks
Sanjay Kumar Gupta · 1993
High transmission speeds, increased burstiness of traffic, and statistical multiplexing of traffic render traditional approaches to network management and control ineffective. This thesis develops the operation and performance of high-speed networks by developing tractable models and approximations. The insight gained is utilized to propose ways of enhancing the efficiency of network resources and facilitating ease of network management and control. Dynamic routing algorithms for routing Virtual Circuits (VCs) in Asynchronous Transfer Mode (ATM) must take into account their heterogeneous bandwidth characteristics and quality of service requirements. We classify ATM networks according to the network characteristics which have the greatest bearing on the performance of dynamic routing algorithms and discuss appropriate routing algorithms for each class and suggest methodologies for performance evaluation. We argue that the problem of designing Virtual Path (VP) subnetworks can be broken down into (i) determining the traffic classes and (ii) designing the overlay, i.e., assigning routes in the physical network to VPs. The problem of determining the optimal set of traffic classes is formulated as a set-partitioning problem. Efficient heuristics that exploit the structure of the problem are proposed. Multi-commodity flow problem is used to determine the overlay. A network that consists of LANs interconnected with leased lines, packet-switched networks, or MANs (or combinations thereof) is referred to as a heterogeneous data network. We develop a performance methodology that accounts for bursty nature of traffic, nested segmentation and reassembly of the packets, window flow control, and bandwidth allocation policies, associated with heterogeneous networks. The performance methodology developed gives good first-cut estimates of end-to-end response time and buffer overflow and requires minimal CPU time. Finally, we examine the Media Access Control (MAC) protocol of FDDI under "overload". We show, for an overloaded system, that the response time of a user that generates packets infrequently can be unacceptably large. Further, the response times grow dramatically with burstiness of traffic. We demonstrate that the response times of "light users" can be improved significantly. In addition, the expected waiting times for heavy users can also be reduced. We propose control mechanisms to force the system to operate in a desirable manner.