Resource Provisioning and Bandwidth Brokering for IP-core Networks Chen-Nee Chuah ISRG Retreat Jan 10-12, 2000 Problem: How to provide end-to-end QoS in IP-core networks in a scalable manner?
ISP2 Example Scenario Resource Reservation ISP1 ISP3 SLA: Agreements that describe the volume of traffic exchanged, bandwidth reserved and price ISP2 SLA H2 H1 H3
Research Issues Resource Provisioning –How to estimate bandwidth usage in advance for capacity planning purposes? Adaptive Reservations –How to adapt aggregate reservations based on traffic fluctuation? –What are the trade-offs between granularity, QoS and signaling complexity? Admission Control –End-to-end? –In stages: Per ISP cloud? Per domain?
ISP2 Hierarchical Clearing House Approach Distributed database –reservation status, % link utilization, optimum path Bandwidth brokering software agent –adapt reservation dynamically source ISP1 ISP n destination Edge Router CH 1 ICH CH 1 CH 2 ISP2
Resource Reservation Strategies Aggregation of reservation requests Hierarchical approach De-couple notifications & reservation requests Static and Dynamic Advanced Reservations Notifications (every u s) - Reservation status - Link utilization - Bandwidth predictor CH 1 ICH CH 2 CH 1 ICH Adapt Reservations - Advance reservations - Process reservation requests ERs aggregate reservation requests (T a )
Traffic Predictors Monitoring system at Edge Router –Online measurement of aggregate rate of incoming & outgoing traffic (regular interval: W est ) Two Traffic predictors for advanced reservations –Local Gaussian predictor for static reservation Larger time-scale (e.g. an hour) Compensate for the coarse granularity of the notifications –Auto-regressive predictor for dynamic reservation Smaller time-scale (W est )
Evaluation Overall Performance Metrics –Link utilization –% blocking/dropping Bandwidth Estimator –How well does the predictor track the traffic fluctuation? –Choice of estimation window, % over-provisioning Signaling between CHs –Sensitivity analysis: effect of aggregation on QoS and complexity Completely de-coupled notifications Limited notifications
Simulation Study: Network Topology vBNS Backbone Network Map (1999) Extreme cases - Dumbbell - Highway with merging paths Houston Seattle SF LA Orlando Atlanta DC NY Denver St. Louise Chicago Boston
Simulation Study: Workload Modeling Two QoS classes –High priority voice calls and video conferencing –Best-effort data traffic (e.g. web, telnet, ftp) Traffic model – Voice & video conferencing calls Poisson arrivals with v and c Erlangs Exponentially distributed call duration (mean = 2.5 min. for voice, 30 min. for video conferencing calls) Individual source is modeled as two state-Markov chain. When “on”, a voice call requires bandwidth of 128 kbps, defined as one basic unit (BU) Video conferencing calls occupy 4 BU –TCP connections get equal share of the non-reserved bandwidth