Architectures for Service Differentiation in Overloaded Internet Servers Architectures for Service Differentiation in Overloaded Internet Servers

Voigt, T. 2002: Architectures for Service Differentiation in Overloaded Internet Servers. SICS Dissertation Series 30. Also as DoCS 02/119. 153 pp. Uppsala. ISBN 91-5061559-9. Web servers become overloaded when one or several server resources such as network interface, CPU and disk become overutilized. Server overload leads to low server throughput and long response times experienced by the clients. Traditional server design includes only marginal or no support for overload protection. This thesis presents the design, implementation and evaluation of architectures that provide overload protection and service differentiation in web servers. During server overload not all requests can be processed in a timely manner. Therefore, it is desirable to perform service differentiation, i.e., to service requests that are regarded as more important than others. Since requests that are eventually discarded also consume resources, admission control should be performed as early as possible in the lifetime of a web transaction. Depending on the workload, some server resources can be overutilized while the demand on other resources is low because certain types of requests utilize one resource more than others. The implementation of admission control in the kernel of the operating system shows that this approach is more efficient and scalable than implementing the same scheme in user space. We also present an admission control architecture that performs admission control based on the current server resource utilization combined with knowledge about resource consumption of requests. Experiments demonstrate more than 40% higher throughput during overload compared to a standard server and several magnitudes lower response times. This thesis also presents novel architectures and implementations of operating system support for predictable service guarantees. The Nemesis operating system provides applications with a guaranteed communication service using the developed TCP/IP implementation and the scheduling of server resources. SILK (Scout in the Linux kernel) is a new networking stack for the Linux operating system that is based on the Scout operating system. Experiments show that SILK enables prioritizing and other forms of service differentiation between network connections while running unmodified Linux applications. Thiemo Voigt, Department of Information Technology, Uppsala University, Box 337, SE-751 05 Uppsala, Sweden, E-mail: [email protected] c © Thiemo Voigt 2002 ISSN 0283-0574 ISSN 1101-1335 ISRN SICS-D--30--SE Printed in Sweden by Elanders Gotab, Stockholm 2002. Distributor: SICS, Box 1263, SE-164 29 Kista, Sweden.