March 8, 2000 1
University of KansasDepartment of Electrical Engineering and Computer Science
Finding Your Shade of Grey on the Network
Spectrum
Towela P.R. Nyirenda-Jere
Victor S. Frost
(sponsored by Sprint)
March 8, 2000 2
University of KansasDepartment of Electrical Engineering and Computer Science
The Network Spectrum
• Simple traffic handling + Huge Capacity
• Moderate traffic handling + Moderate Capacity
• Complex traffic handling + Minimal Capacity
March 8, 2000 3
University of KansasDepartment of Electrical Engineering and Computer Science
The Problem
• Determining the equivalence of traffic handling mechanisms
• Understanding the trade-off between the complexity of traffic handling mechanisms and the network capacity required to support service guarantees
March 8, 2000 4
University of KansasDepartment of Electrical Engineering and Computer Science
What we know
• Service guarantees depend on both traffic handling and network capacity
• Total aggregation schemes require more capacity than per flow schemes
• Partial aggregation schemes scale better than per-flow schemes when number of flows is large
March 8, 2000 5
University of KansasDepartment of Electrical Engineering and Computer Science
What we don’t know
• How much more capacity do we need with total aggregation versus per-flow?
• How does the complexity of per-flow management measure up against the cost of additional capacity with aggregate traffic handling?
• What about partial aggregation schemes?
March 8, 2000 6
University of KansasDepartment of Electrical Engineering and Computer Science
What we need
• Quantification of the gain obtained by using complex traffic handling with smaller network capacity versus using simple traffic handling with abundant network capacity
• Quantification of the sensitivity of traffic handling to changes in network load both in terms of the total load and in terms of the relative mix of different classes
March 8, 2000 7
University of KansasDepartment of Electrical Engineering and Computer Science
What should be done
• Quantify trade-off between complexity of traffic handling and network capacity
• Determine scalability of the analytic methods with network size and capacity
• Provide analytic framework for capacity provisioning and traffic handling strategy
• Study the sensitivity of traffic handling schemes to changes in network load
March 8, 2000 8
University of KansasDepartment of Electrical Engineering and Computer Science
Traffic Handling Mechanisms
March 8, 2000 9
University of KansasDepartment of Electrical Engineering and Computer Science
Traffic Modeling
• Deterministic burstiness constraint model of Cruz et. al.
• Traffic described by two parameters: average rate and burstiness
• No assumptions on traffic type
• Aligns well with IETF and ATM Forum traffic description
March 8, 2000 10
University of KansasDepartment of Electrical Engineering and Computer Science
Analytic Method
• Use Network Calculus approach of Cruz, Parekh & Gallagher
• WFQ is used as the reference mechanism
• Find number of voice, video, e-mail and WWW sources using WFQ taking into account delay requirements
• Find capacity required to support these sources using CBQ, PQ and FIFO
March 8, 2000 11
University of KansasDepartment of Electrical Engineering and Computer Science
Applications and Service Requirements
March 8, 2000 12
University of KansasDepartment of Electrical Engineering and Computer Science
Scenario
• Single-node Network
• OC-3 Link for WFQ
• Video load = 10% of OC-3
• Voice load varied from 10-90% of OC-3
• E-mail and WWW share remaining capacity using pre-defined ratios
March 8, 2000 13
University of KansasDepartment of Electrical Engineering and Computer Science
CBQ Capacity Requirements• Capacity
requirements of CBQ same order of magnitude as WFQ
March 8, 2000 14
University of KansasDepartment of Electrical Engineering and Computer Science
PQ Capacity Requirements
• Capacity requirements of PQ same order of magnitude as WFQ
• Non-monotonic
March 8, 2000 15
University of KansasDepartment of Electrical Engineering and Computer Science
PQ and CBQ Capacity Requirements
• PQ capacity does not exceed CBQ capacity
March 8, 2000 16
University of KansasDepartment of Electrical Engineering and Computer Science
FIFO Capacity Requirements• FIFO requires
two orders of magnitude more capacity than WFQ
March 8, 2000 17
University of KansasDepartment of Electrical Engineering and Computer Science
Sensitivity to Design Point
• Goal is to explore the ability of the three schemes to provide acceptable delay guarantees when the traffic submitted exceeds the traffic for which the network was designed
• Two broad cases» voice as dominant class» WWW as dominant class
March 8, 2000 18
University of KansasDepartment of Electrical Engineering and Computer Science
Sensitivity: Network designed for Voice
• WFQ1, CBQ sensitive to increase in voice
• PQ, FIFO not as sensitive
March 8, 2000 19
University of KansasDepartment of Electrical Engineering and Computer Science
Sensitivity: Network designed for WWW
• FIFO most sensitive to increase in WWW traffic
• PQ least sensitive
March 8, 2000 20
University of KansasDepartment of Electrical Engineering and Computer Science
Projections on Network Traffic and Capacity
• Assume 5% growth in Voice and 100% growth in WWW per year
• Initially OC-3 link with total utilization 45%» 5% voice, 15% e-mail and 25% WWW
March 8, 2000 21
University of KansasDepartment of Electrical Engineering and Computer Science
Projections on Network Traffic and Capacity
• FIFO capacity at year 5 is 2000x capacity at year 1
March 8, 2000 22
University of KansasDepartment of Electrical Engineering and Computer Science
Projections on Network Traffic and Capacity
• CBQ and PQ capacity at year 5 is 8x capacity at year 1
• WFQ capacity at year 5 is 4x capacity at year 1
March 8, 2000 23
University of KansasDepartment of Electrical Engineering and Computer Science
Shades of Grey
March 8, 2000 24
University of KansasDepartment of Electrical Engineering and Computer Science
What Have We Learned
• It is possible to quantify the trade-off between network capacity and traffic management
• Sensitivity of the traffic handling schemes depend on the assumptions made in designing the network as well as the traffic class contributing to the growth in traffic
March 8, 2000 25
University of KansasDepartment of Electrical Engineering and Computer Science
What’s Next
• Review methodology and define performance metrics/indices
• Extend analysis to carrier-size networks
• Incorporate stochastic bounds on performance
March 8, 2000 26
University of KansasDepartment of Electrical Engineering and Computer Science
Significance
• Identifying the tradeoffs associated with the use of traffic handling mechanisms with respect to network capacity
• Sensitivity analysis will provide a tool for long-term planning
• Development of a methodology which can be used to compare traffic handling schemes in general