Predictive End-to-End Reservations via A Hierarchical Clearing House Endeavour Retreat June 19-21, 2000 Chen-Nee Chuah (Advisor: Professor Randy H. Katz) EECS Department, U. C. Berkeley
Problem Statement How to deliver end-to-end QoS for real-time applications over IP-networks? Video conferencing, Distance learning Web surfing, s, TCP connections Internet PSTN VoIP (e.g. Netmeeting) H.323 Gateway GSM Wireless Phones
Why Is It Hard? Lack of QoS assurance in current IP-networks –SLAs are not precise Scalability issues Limited understanding on control/policy framework –How to regulate resource provisioning across multiple domains? ISP1 ISP 3 ISP2 H3 ?? SLA H1
Example Workload: Real-Time Packet Audio Wide range of audio intensive applications –Multicast lecture, video conferencing, etc. –Significantly different from 2-way conversations –Traffic characteristics too diverse, cannot be described by one model Resource pre-partitioning doesn’t work! Application Specific Traffic Patterns
Proposed Solution: Predictive Reservations H1 H2 LCH Edge Router Online measurement of aggregate traffic statistics Advance reservations based on local Gaussian predictor –R A = m + Q -1 (p loss ). Allow local admission control Advance Reservation Dynamic Reservation
Predictor Characteristics 1-min predictor % Loss - 7 % Over-Prov. 10-min predictor - 0.7% Loss - 33 % Over-Prov. More BW for BE traffic than pre-partitioning - avg. 286 Kbps - max Kbps
Reservations Across Multiple Domains via A Clearing House Architecture Introduce logical hierarchy Distributed database –CH-nodes maintain reservation status, link utilization, network performance source ISP1 ISP n destination Edge Router LCH CH 2 ISP2 CH 1
Clearing House Approach Delivers statistical QoS –Aggregate reservation requests –Coordinates aggregate reservations across multiple domains –Performs coarse-grained admission control in a hierarchical manner Assumptions –Networks can support differentiated service levels –Traffic and network statistics are easily available Independent monitoring system or ISPs –Control and data paths are separate
Advantages Maintain scalability by aggregating requests –Core routers only maintain coarse-grained network state information Provide statistical end-to end QoS –Advance reservations & admission control Reduce setup time –Advance reservations allow fast admission control decisions Optimize resource utilization –Predictive reservations achieve loss rate < 1% without extensive over-provisioning
Future Work: Simulation Study vBNS backbone network topology (1999) Houston Seattle SF LA Orlando Atlanta DC NY Denver St. Louise Chicago Boston !Traffic matrix weighted by population !Three-level Clearing House architecture - one top CH-node - one CH-node per city - local hierarchy of LCHs Workload models: two QoS classes –High priority packet audio 25 traces (conference & telephone calls), minutes –Best-effort data traffic