Quality of Service Performance of ATM and IP.

Abstract

Asynchronous Transfer Mode (ATM) with its Constant Bit Rate (CBR) and Variable Bit Rate (VBR) services is a network architecture tailored to real-time applications. The Internet Protocol (IP) systems--though initially designed for non-real-time data--are now being augmented with Integrated Services IP (IntServ) and Differentiated Services IP (DiffServ) in order to support these applications. Given the network circumstances, what Quality of Service (QoS) should the real-time applications expect from the above services? Can these services commit to the requested QoS of the real-time applications? This thesis attempts to answer the above questions by evaluating and modeling the performance of the underlying ATM and IP layers. We study the phase dependency in the CBR delay process, and propose a Gamma distribution to model this process. The generality of this model over variety of ATM network scenarios is tested. We analyze and formulate the Cell Delay Variation (CDV) of CBR and VBR connections. We demonstrate that the receiver-time can be divided into several epochs. We find a periodic behavior in the cell-delay-correlation of a CBR connection when being multiplexed with similar (homogeneous) connections. Next, we consider a heterogeneous environment consisting of ATM and IP networks. Resources at both ATM and IP nodes (switches) are reserved by the Guaranteed Service IP, one of the two services comprising IntServ. The essential factors in this investigation are the overhead of various protocol layers and the size of the packets generated by the application. Insofar as packet delay-jitter is concerned, the performance of IP-over-ATM is very close to that of the native ATM. Although, the buffer size requirement of the native ATM is considerably less, IP-over-ATM can still be superior because of the low memory cost. We implement the DiffServ framework comprised of Premium, Assured, and Best effort service classes. We demonstrate the superiority of Diffserv over the per-flow services. We investigate the ability of the RIO ('Random Early Discard with In and Out') mechanism to effectively discriminate the Assured traffic from the multiplexing best-effort traffic. We dynamically measure the traffic existing inside the network, and apply the measurement to the network control mechanisms. We examine the Weighted Fair Queuing (WFQ) mechanism with dynamic weight assignment. We find that WFQ can only marginally improve the network performance. Finally, we show that shaping a Premium flow can exceedingly increase the overall packet delay and jitter. (Abstract shortened by UMI.).