Scheduling Heterogeneous RealTime Traffic over Fading Wireless Channels I-Hong Hou P.R. Kumar University of Illinois, Urbana-Champaign 1/24 Background: Wireless Networks There will be increasing use of wireless networks for serving traffic with QoS constraints: Example: VoIP, Video Streaming, Real-time Monitoring, Networked Control, etc. Client requirements include Specified traffic patterns Delay bounds Timely throughput bounds 2 /24 Previous Work and Challenges Prior work [Hou et al] [Hou and Kumar]: Q: How to deal with more complicated scenarios? Clients have hard throughput requirements Static but unreliable wireless channels All clients require the same delay bounds Optimal packet scheduling policies are proposed Rate adaptation may be applied Channel qualities can be time-varying Clients may require different delay bounds This work extends the model in prior work and proposes a guideline for these scenarios 3 /24 Client-Server Model A system with N wireless clients and one AP Time is slotted AP schedules all transmissions 2 1 AP 3 4 /24 More General Traffic Model {1,.,3} Group time slots into periods with T time slots Clients may generate packets at the beginning of each period {.,2,.} {1,2,3} {1,.,3} {.,2,.} {1,2,3} T 2 1 AP {1,.,3} 3 {.,2,.} {1,2,3} 5 /24 Different Delay Bounds arrival Deadline for client n = τn τ1=4 2 1 deadline AP 3 τ3=3 arrival deadline τ2=5 arrival deadline 6 /24 Channel Model Channel changes from period to period Channels are static within a period System may or may not support rate adaptation With Rate Adaptation Transmission takes sc,n time slots under channel c Transmissions are error-free Without Rate Adaptation Transmission takes 1 time slot Transmissions succeeds with probability pc,n 7 /24 Timely Throughput Requirements Timely throughput # of delivered packets = # of periods Client n requires timely throughput qn Q: How to design a scheduling policy to fulfill requirements of all feasible sets of clients? Feasibility optimal scheduling policy 8 /24 Pseudo-debt Delivery debt: deficiency of timely throughput (t / T )qn # of packets delivered for client n Time debt: deficiency of time spent on a client Pseudo-debt rn(t) quantifies the behavior of client n up to time t rn (t ) ] converges The set of clients is fulfilled [ t to 0 in probability 9 /24 Sufficient Condition for Optimality Let μn be the reduction on debt for client n Theorem: A policy that maximizes E{ rn (t ) n } for each period n is feasibility optimal. Analogous to Max-Weight scheduling in wireline networks 10 /24 Rate Adaptation with Different Delay Bounds Scenario: Rate adaptation used Clients may have different per packet delay bounds, τn Modified Knapsack Policy: Find an ordered set S={m1,m2,…} to maximize total debt τ1=4 τ2=7 τ3=10 S1 = 3 S2 = 5 S1 = 3 S3 = 4 S3 = 4 A variation of knapsack problem and can be solved by DP 11 /24 Rate Adaptation with Different Delay Bounds Scenario: Rate adaptation used Clients may have different per packet delay bounds, τn Modified Knapsack Policy: Find an ordered set S={m1,m2,…} to maximize total debt τ1=4 τ2=7 τ3=10 S1 = 3 S2 = 5 S2 = 5 S3 = 4 S3 = 4 A variation of knapsack problem and can be solved by DP 12 /24 Rate Adaptation with Different Delay Bounds Scenario: Rate adaptation used Clients may have different per packet delay bounds, τn Modified Knapsack Policy: Find an ordered set S={m1,m2,…} to maximize total debt τ1=4 τ2=7 τ3=10 S1 = 3 S2 = 5 S1 = 3 S2 = 5 S3 = 4 A variation of knapsack problem and can be solved by DP 13 /24 Time-Varying Channels Scenario: Joint Debt-Channel Policy: Same delay bounds for all clients, τ≡τn Time-varying channels, pn(t) Applicable to Gilbert-Elliot fading Model Let rn(t) be delivery debt Clients with larger rn(t) pn(t) get higher priorities Theorem: The Joint Debt-Channel policy is feasibility optimal 14 /24 Heterogeneous Delay Bounds Scenario: Static channels, pn≡pn(t) Different delay bounds for all clients, τn Adaptive-Allocation Policy: Let rn(t) be time debt Estimate the # of slots needed by client n for a successful transmission, ηn Dynamically allocate slots to maximize rn (t ) n n 15 /24 Evaluation Methodology Evaluate four policies: Proposed policies for each scenario PCF with randomly assigned priorities (random) Two policies proposed by [Hou, Borkar, and Kumar] Time debt first policy Weighted-delivery debt first policy Metric: Total delivery debt 16 /24 Rate Adaptation: VoIP Setup Period length = 20 ms Two groups of clients: Group A One packet every 60 ms Group B One packet every 40 ms 21.3 kb/s traffic require 19.2 kb/s timely throughput Starting times evenly spaced 32 kb/s traffic require 22.4 kb/s timely throughput Data rates alternate between 11 Mb/s and 5.5 Mb/s 66 Group A clients and 44 Group B clients 17 /24 Rate Adaptation: VoIP Results 18 /24 Time-Varying Channels: VoIP Setup Period length = 20 ms Two groups of clients: Group A One packet every 60 ms Group B One packet every 40 ms 21.3 kb/s traffic require 19.2 kb/s timely throughput Starting times evenly spaced 32 kb/s traffic require 22.4 kb/s timely throughput Channel evolves based on Gilbert-Elliot model 57 Group A clients and 38 Group B clients 19 /24 Time-Varying Channels: VoIP Result 20 /24 Heterogeneous Delay Bounds: VoIP Setup Two groups of clients: Group A One packet every 60 ms 21.3 kb/s traffic Group B One packet every 40 ms 32 kb/s traffic require 19.2 kb/s timely throughput Delay bound = 20 ms Starting times evenly spaced require 22.4 kb/s timely throughput Delay bound = 13 ms Average channel reliabilities between 80% and 96% 57 Group A clients and 38 Group B clients 21 /24 Heterogeneous Delay Bounds: VoIP Result 22 /24 Conclusion Extend previous model for more complicated scenarios With or without rate adaptation Time-varying channels Heterogeneous delay bounds Identify a sufficient condition for optimal scheduling policies Design policies for several cases Time-varying channels, heterogeneous delay bounds with rate adaptation Time-varying channels without rate adaptation Heterogeneous delay bounds without rate adaptation 23 /24 24/24 Rate Adaptation: MPEG Setup Period length = 6 ms Two groups of clients: Group A Group B 1700 kb/s traffic 1360 kb/s traffic require 1530 kb/s timely require 816 kb/s timely throughput throughput Data rates alternate between 54 Mb/s and 24 Mb/s 6 Group A clients and 6 Group B clients 25 Rate Adaptation: MPEG Results 26 Time-Varying Channels: MPEG Setup Period length = 6 ms Two groups of clients: Group A Group B 1700 kb/s traffic 1360 kb/s traffic require 1530 kb/s timely require 816 kb/s timely throughput throughput Average channel reliabilities between 80% and 89% 4 Group A clients and 4 Group B clients 27 Time-Varying Channels: MPEG Setup 28