We propose a quality of service (QoS) architecture, QUANTA, for an end system protocol suite. We use TCP(UDP)/IP over ATM as a testbed to develop the architecture. We measure the application level QoS in terms of throughput, delay, round trip time, and loss to identify the base line performance an application can expect from such an environment. From the no-load condition we measure the behavior of these protocols at various data rates and user submitted data block sizes. We demonstrate the tradeoffs involved in obtaining high throughput, low delays, low round trip time, and zero losses at different data rates. We use host load condition experiments to understand the interaction between the CPU intensive jobs and the communication intensive jobs. We use network load condition experiments to observe interaction between multiple streams of the above two protocol suites, and its effect on the application QoS. Given these observations we define the missing components in the current protocol architectures to provide tighter control on the QoS guarantees. Components we define in QUANTA include a two level application to network QoS translator protocol tuning components, local feedback component, class based scheduling etc.

The Distributed Queue Dual Bus (DQDB) system consists of a linear arrangement of N nodes that communicate with each other using two contra-flowing buses. The nodes use an extremely simple protocol to send messages on these buses. This simple, but elegant, system has been found to be very challenging to analyze. We consider a simple and uniform abstraction of this model to highlight the fairness issues in terms of average waiting time. We introduce a new approximation method to analyze the performance of DQDB system in terms of the average waiting time of a node expressed as a function of its position. Our approach abstracts the intimate relationship between the load of the system and its fairness characteristics, and explains all basic behavior profiles of DQDB observed in previous simulation. For the uniform DQDB with equal distance between adjacent nodes, we show that the system operates under three basic behavior profiles and a finite number of their combinations that depend on the load of the network. Consequently, the system is not fair at any load in terms of the average waiting times. We also show that the main theme of the analysis carries over to the general (nonuniform) DQDB. By suitably choosing the inter-node distances, the DQDB can be made fair around some loads, but such system will become unfair as the load changes. In the vicinity of a critical load, the uniform network runs into a state of instability, where its behavior fluctuates from one extreme to the other with small load variations. Our analysis is supported by simulation results.>

Considers a simple and uniform abstraction of the distributed queue dual bus (DQDB) system of N nodes to highlight the fairness issues in terms of average waiting time. For the uniform DQDB with equal distance between adjacent nodes, the authors show that the system operates under three basic behavior profiles and a finite number of their combinations that depend on the load of the network. Consequently, the system is not fair at any load in terms of the average waiting times. In the vicinity of a critical load of 1-4/N the uniform network runs into a state akin to chaos, where its behavior fluctuates from one extreme to the other with a load variation of 2/N. The analysis is supported by simulation results. The authors also show that the main theme of the analysis carries over to the general (non-uniform) DQDB.>