PRISM: Proxies for Internet Streaming Media J. Kurose, P. Shenoy, D. Towsley (UMass/Amherst) L. Gao (Smith College) G. Hjalmtysson, J. Rexford (AT&T Research - Labs)

Outline o goals of project o services for streaming media o research issues o progress to date

Vision: enabling/enhancing a new service infrastructure First generation servers o single server, multiple clients o end-end principle o homogeneity in offered service 2nd generation service: proxy between server, clients client-tailored services enhanced scalability

Proxy Services: Motivation Improved performance o protect user from poor throughput, delay, loss o exploit short RTT to client o exploit single provider of proxy-client path Value-added services o services not available at server, tailored to client, transcoding, error recovery, interactive services Exploit inside information o knowledge of QoS/provisioning on proxy-client path o knowledge of client bandwidth and buffer resources

Research Goals o Identification of proxy services o System research issues, operating systems for proxies: resource allocation for QoS, middleware/interfaces for programmable services, network protocols: error recovery, prefetch/prefix caching, congestion control, service location/invocation o analysis, simulation, prototype and experiment

Sample Proxy Services o patching o quality enhancement o transcoding

Patching Service o problem, asynchronous requests for long video streams, low playback delay requirement o basic idea, client receives ongoing transmission from server, retrieves initial portion from proxy video flow request

Multiple Clients o Problem: multiple asynchronous requests for video o basic ideas, batching, patching, segmented delivery, caching, schedule, multicast video segments over different multicast groups video flow

Quality Enhancement o problem, lossy, time-varying bw path from server to proxy o basic idea, proxy performs error recovery (ARQ, FEC) congestion control/buffering with server, simple clients see enhanced flow enhanced video flow video data feedback

Operating System Support o flexible resource allocation mechanisms in proxy OS o rate-based CPU, network, and disk schedulers, Challenge: design rate-based allocation mechanisms for large SMPs, disk arrays and multi-homed hosts o resource reservation mechanisms that provide performance guarantees to applications o Storage management techniques for proxy workloads, Challenge: design techniques to handle workloads that frequently read, write and delete diverse objects from disk cache o Platform: Use QLinux, a QoS-enhanced version of Linux that we have developed

Middleware/Interfaces: Pronto o programmable, active services o separation of control, forwarding o control:, out of data path, best effort, asynchronous o implementation:, stream thinning, active congestion control, reliable mcast, traffic shaping data path forwarding engine interface Meta control

Network Protocols Client/server/proxy o batching, patching, segmented delivery, catching protocols, data transfer: use of multiple multicast channels, control, signaling o Congestion control, thinning, discard, separate server-proxy, proxy-client “connections” Reliability enhancement:, active repair services, FEC, time-critical error control

Current Status o server/client implementation, addition of proxy, services o QLinux: current version based on 2.2.x kernel, rate-based CPU, packet, and disk schedulers o PRONTO o partner with Z. Zhang (UMinn), distribution of software

Summary o Proxy services, enhanced performance, valued added services, proxy, clients “near” each other o System research issues, operating systems resource allocation for QoS, middleware for programmable services, network protocols o Analysis, simulation, prototype and experiment