¡°Speculative Data Dissemination and Service¡±

	This paper proposes 2 server-initiated protocols to improve the performance of a distributed information system: 1) propagate information from producer to servers that are closer to consumer spatially; 2) speculate future requests based on current and send ¡°extra¡± data with current request. The first protocal is based off the observation that 0.5% of all available documents account for 69% of all requests; and 10% of all documents account for 91% of all requests. Simulation shows that 20 proxies, each of size 100 MB, could save up to 35% of bandwidth usage. The second protocol reduces server load by sending speculated future requests at the cost of increased traffic. By using 5% extra bandwidth, a 30% reduction in server load is observed. However, this approach has a high diminishing return.

	I do not find the paper to be very practical. Both protocols are aimed at reducing bandwidth and server load. However, both protocols has very high costs associated. In the first protocol, 20 proxies could save 35% of bandwidth. However, bandwidth is a lot cheaper than hardware. The cost of 20 proxies could easily cover the 35% extra bandwidth used. Furthermore, 20 proxies would only reduce the bandwidth of 1 website. Increasing the capacity of proxies to handle multiple websites would incur more significant costs. The second protocol relies heavily on proper speculation logic and client access patterns. In order for the speculation protocol to be viable, it must not incur significant server overhead. Furthermore, the speculation protocol seems more exploitable to denial of service attacks.