Handling Persistent Connections in Overloaded Web Servers

Web servers have to be protected from overload since overload can lead to unpredictable response times, low throughput and even loss of service. To protect servers from overload several admission control archi-tectures have been developed [1, 4, 5, 6, 2, 3, 8]. Many of these architectures can reject or accept connection requests based on information provided in the HTTP header [3, 8]. Persistent connections allow HTTP clients to send several requests on the same TCP connection to reduce client latency and server overhead [7]. Persistent connections are also a challenging problem with respect to admission control since the resource consumption of the requests is unknown at the time the admission control decision has to be made, i.e. when the web server receives the first request on a persistent connection. For example, a user visiting the home page of a company might either leave the company’s web pages after visiting a single page or might initiate a long session. Obviously, a long session demands much more resources. Admission control is thus a trade-off. If admission policies are too restrictive, potential customers might be rejected unnecessarily resulting in loss of revenue. If one is too optimistic, the server may become overloaded with unpredictable response times and low throughput as a possible consequence. Even well-engineered adaptive overload prevention schemes suffer from this problem. For example the onset of overload may not be predicted with sufficient accuracy due to workloa