This thesis investigates TCP connection management mechanisms in order to understand the behaviour and improve the performance of Internet servers during overload conditions such as flash crowds. We study several alternatives for implementing TCP connection establishment, reviewing approaches taken by existing TCP stacks as well as proposing new mechanisms to improve server throughput and reduce client response times under overload. We implement some of these connection establishment mechanisms in the Linux TCP stack and evaluate their performance in a variety of environments. We also evaluate the cost of supporting half-closed connections at the server and assess the impact of an abortive release of connections by clients on the throughput of an overloaded server. Our evaluation demonstrates that connection establishment mechanisms that eliminate the TCP-level retransmission of connection attempts by clients increase server throughput by up to 40% and reduce client response times by two orders of magnitude. Connection termination mechanisms that preclude support for half-closed connections additionally improve server throughput by up to 18%.

This thesis investigates TCP connection management mechanisms in order to understand the behaviour and improve the performance of Internet servers during overload conditions such as flash crowds. We study several alternatives for implementing TCP connection establishment, reviewing approaches taken by existing TCP stacks as well as proposing new mechanisms to improve server throughput and reduce client response times under overload. We implement some of these connection establishment mechanisms in the Linux TCP stack and evaluate their performance in a variety of environments. We also evaluate the cost of supporting half-closed connections at the server and assess the impact of an abortive release of connections by clients on the throughput of an overloaded server. Our evaluation demonstrates that connection establishment mechanisms that eliminate the TCP-level retransmission of connection attempts by clients increase server throughput by up to 40% and reduce client response times by two orders of magnitude. Connection termination mechanisms that preclude support for half-closed connections additionally improve server throughput by up to 18%.

Providing service via theWeb differs from other service provision environments in that it is possible for the unexpected arrival of a massive number of service requests in a small time-frame, a situation commonly referred to as a flash crowd. Events of this nature are beyond the control of the service provider, and have the potential to severely degrade service quality and, in the worst case, to deny service to all clients completely. The occurrence, severity and sought Web content of a flash crowd is beyond the control of service provision software. How this software reacts to such a flash crowd, however, is not. Given the short-lived nature of flash crowds, it is unreasonable to expect such systems to increase the system resources they can apply to a particular flash crowd event. It is also difficult to predict the particular nature of any flash crowd, and subsequently which system resources will bottleneck. The driving hypothesis of this research is that, if we are to reasonably expect to have software react effectively to flash crowd events, we need to alter that software at runtime to remove system bottlenecks, whilst a flash crowd event is in progress. This is a special case of what is usually known as "unanticipated software evolution". This thesis reports on an investigation into how unanticipated software evolution can be applied to running Web service provision software to remove system bottlenecks. It does so by introducing automated dynamic Web content degradation to running software currently subject to simulated flash crowd events. The thesis describes and validates appropriate runtime extensions to allow generative object communication architectures (a promising class of architecture for unanticipated software evolution) to be converted initially into a Web application server, and then later accept further runtime behaviour changes. Such changes could alter system bottlenecks by replacing the key programming logic causing system bottlenecks at runtime.