Travis Portz
Large, sudden increases in the traffic to a website Low-traffic website being linked to by a popular news feed “Slashdot Effect” Popular news sites receiving increased traffic due to a major world event Effects of flash crowd Reduced performance Website inaccessible
Content distribution network (CDN) Embedded objects stored on external servers Commercial solutions are expensive for low-traffic websites Drop requests Not a real solution Proposed alternative: Web servers collaborate in a peer-to-peer content distribution network
Estimate request arrival rate λ for each object on the web server Exponential moving average Similar to TCP’s estimation of RTT
CONTENT REPLICATION If arrival rate exceeds λ R, start replicating content on peer servers Make one copy for each request received Stop when all peers have copies REQUEST FORWARDING If arrival rate exceeds λ F, start forwarding requests to peer servers HTTP redirection Round-robin order
Not all requests are forwarded once arrival rate exceeds λ F Origin server handles some requests, forwards other requests with probability P F Probability increases with Request arrival rate Time since last request forwarded to peer server
Inputs List of servers and structure of overlay network Baseline request rates Request rates for flash crowd objects Server bandwidth and processing limits
Outputs Actual server request rates Request drop rates Bandwidth usage