High Performance,Scalable Web Servers

Graduate

Our project addresses problems related to the development of high performance,scalable web server system. Our achievement can be divided into two parts: single web server and multiple replicated web servers. For single web server, Browser Initiated Pushing(BIP) is proposed to improve performance based on the observation that today's typical web page has one or more embedded images. BIP reduces the user perceived response time by 40% under normal load and 3-4 times under heavy load. Also, BIP increases server's throughput by 30% and reduces the overhead of traditional pushing from 32% to 10%. For replicated web servers, we proposed a non-dispatcher approach Static Allocation and Client Redirection (SSCR) that shares load between replicated web servers and works with both clusters of workstations and web server replication at global scale. SSCR offers competitive performance to dispatcher-based approach without one-point bottleneck and failure. Smart Server Selection(S3) addresses the important problem: when web servers are replicated at global scale, how can the user automatically access nearest server available according to network metrics? In S3, client DNS server is extended to prioritize a pool of IP addresses based on the routing metric information collected from routers and other information it collects ( geographical location of servers and clients). An efficient scheme to collect routing-metric information from routers is proposed. The overhead of such an scheme is independent of number of replicated servers. This is a big achievement over existing approaches because the overhead of these approaches suffers a linear increment as the number of replicated web servers increase. Ongoing efforts include supporting differentiated web services under replicated web servers, effective management for network services in general, and replicated web servers in particular.

Return to Workshop Listing