This proposal aims to provide this deep understanding of server behavior during overload, in the case of both static and dynamic content. The approach is interdisciplinary: There is a theoretical component which aims to prove new theorems on M/G/1 queues with time-dependent load; an implementation component which performs a rigorous systematic study of web server and database server performance under overload as a function of various networking parameters; and a workload and client characterization component which aims to understand statistical properties of the request stream during transient overload and incorporate these into the design of a web workload generator for overload. This proposal also introduces a very simple idea for improving the performance of web servers under overload: a kernel-based implementation of SRPT connection scheduling. The idea is simple and intuitive. Connection scheduling does not require purchasing more hardware, or any other costly system upgrades. It appears to allow the server to survive much longer periods of overload, and to allow clients to experience better performance during these periods of overload. Furthermore, it appears to enable the server to recuperate from overload much faster. The proposal will provide a detailed evaluation of the benefits of SRPT connection scheduling and will aim to introduce similar scheduling policies for databases. The above concludes the intellectual merit portion of the proposal. The broader impacts resulting from the proposed activity include online availability of all source code generated during the project including modifications to the Apache code and to Linux, as well as online availability of all trace logs collected during this project, and a web workload generator for overload. The project initiates collaborations across various academic disciplines including: mathematics, computer systems, and operations management, as well as between universities and industry. The project will also provide a research opportunity for a minority undergraduate student. The proposal involves 1 faculty member (1 summer month only), 1.5 graduate students, and 1 undergraduate, over the course of 3 years.
