Quality of service modeling for wide area network based systems
Paritosh Dixit, Mladen Alan Vouk, Sastry G. Pantula · Medical Entomology and Zoology · 1998
This work addresses the problem of statistical modeling and management of user level Quality of Service (QoS) for wide area network-based systems. QoS is traditionally evaluated using metrics collected at the network layer. However, emerging high performance interactive applications require support guarantees at the application layer, and provision for QoS requirements of individual users regardless of the performance offered by underlying services. In this dissertation, QoS is addressed using two non-traditional approaches: user perceived system and application reliability and availability, and user perceived application response times. A wide area network-based education system called NovaNET was used to study the reliability and availability issues. Internet I (current incarnation of the Internet) and the Web were used to assess user level response time issues. It is argued that any violation of the QoS thresholds, whether at the application level or in the underlying infrastructure can be considered as a user perceived failure of the system to deliver its QoS promise unless the failure is mitigated, or masked, from the user. In other words, reliability and availability of the services must be considered as part of the overall QoS model. Ten years of NovaNET performance data were analyzed to assess reliability and availability parameters that are important for these types of distributed applications. It is shown that relatively simple statistical models, such as the exponential failure-decay model, can be used to adequately describe and predict the reliability and availability of such a system. But, it is also shown that a full understanding of the user level reliability and availability issues can only be obtained through a thorough understanding of more complex failure phenomena, such as the correlation that might occur among system failure states due to common cause faults. Based on NovaNET performance information, user satisfaction with educational systems requires end-user availability levels in excess of 0.95 and keystroke round trip response times of less than about 250 milliseconds, on the average. Experimental measurements of round trip delay times and application response times using the current incarnation of the Internet, Web servers and Web browsers show that (a) there are strong cyclic components in these response times that derive from cyclic loads on the network, (b) that there is considerable variability in the round trip response times that one can expect of Internet-based applications available today and, (c) for most part of the working day, the QoS offered to a typical user, as expressed through round trip delay time, violates the 250 millisecond delay bound and makes real-time interactions very difficult across larger Internet distances. A series of models were developed based on statistical analyses of the experimental data. They range from simple mean-value models to stationary ARMA (p,q) time series models. It is shown that a combination of the two can describe well both the long term and the short term behavior of Internet delays as perceived by a user. The models were validated, for predictions ranging from thirty minutes into the future to one week into the future. An architecture and a set of mechanisms that use the above metrics and models to assess and control end-user level QoS are described. It is assumed that the feedback, and both proactive and reactive control, is available from the application and the networking sides, of the QoS API (application programming interface).