Network-oriented controls of Internet services

Network-Oriented Service Controls (NOSCs) are light-weight mechanisms for providing Quality-of-Service (QoS) support and overload protection for Internet services. They operate purely at the network and transport layers by policing incoming requests. Thus, key to implementing effective NOSCs is understanding how arriving requests react to the applied control and how services react to the change in incoming load. This dissertation undertakes a two-sided investigation of NOSCs focusing on the underlying traffic and the running services to give a complete picture detailing the performance, effectiveness, and limitations of such mechanisms.; Based on a detailed study of the underlying network traffic, we introduce persistent clients, a model that captures the true reaction of client requests to the applied controls. This new model has strong implications on the controllability of aggregate request arrivals. It is used to show the limitations of traditional rate-based mechanisms as they unnecessarily exacerbate client-perceived latency as well as the number of retransmissions.; We propose persistent dropping (PD), a novel mechanism that minimizes both the client-perceived latency and the number of retransmissions without sacrificing fairness. To further improve the efficiency of PD, we propose the Abacus Filter ( AF), an adaptive filtering mechanism that is equipped with an explicit reject capability. AF only drops those requests that, with high probability, will be accepted in the future upon their retransmission and rejects those requests that will never be accepted, further minimizing the client-perceived latency.; We also study the impact of NOSCs on Internet services with particular focus on (1) the effects of concurrency on the performance of a service, (2) the interactions between threads/processes of different services, and (3) the relationship between request arrival and service performance. We use our derived models to predict the impact of the applied (rate-limiting) controls on the client-perceived latency. We also use them to find the enforcement rates that guarantee certain response times to different client groups.; We finally create an adaptive architecture called Adaptive Packet Filters (APFs). By incorporating on-line monitoring information, APFs reduce incoming service requests to provide overload protection and QoS differentiation for different client groups without modifying the application or the underlying OS.