Presentation on theme: "Opportunistic Scheduling Algorithms for Wireless Networks"— Presentation transcript:
1 Opportunistic Scheduling Algorithms for Wireless Networks Vegard HasselCUBAN Seminar 22. April 2004
2 Agenda What is scheduling/opportunistic scheduling? Cross-layer design Different types of channel modelsFairnessAlgorithms that only consider the channel conditionsAlgorithms that consider QoSAlgorithms that consider power consumptionCurrent & future research
3 What Is Scheduling?Scheduling policy:-a rule that specifies which user is allowed to transmit and which user is allowed to receive at each timeslotUplink (user transmits) and downlink (user receives) at different frequencies.
4 What Is Opportunistic Scheduling? Opportunistic: Scheduler tries to exploit channel conditions to achieve higher network performanceSCHEDULERUSER 1USER 2BUFFERSUSER 3The base station serves as a scheduling agent
5 Qualcomm Example (1)The channel conditions for each user is independentThe channel is GOOD 50% of the time and BAD 50% of the timeTwo users with bitrates:User 1: 200Mbit/s or 400Mbit/sUser 2: 400Mbit/s or 800Mbit/sThe users cannot be active at the same timeRound robin algorithm without opportunistic scheduling:R1=0.5*(0.5*200Mbit/s+0.5*400Mbit/s)=150Mbit/sR2=0.5*(0.5*400Mbit/s+0.5*800Mbit/s)=300Mbit/s
6 Qualcomm Example (2) But what if both users need 200Mbit/s? With opportunistic scheduling, the user that has the GOOD channel condition is chosen:25% of the time both users have BAD channel50% of the time one user has GOOD channel25% of the time both users have GOOD channelThe user with the relatively best channel is chosenRound robin with opportunistic scheduling:R1=0.5*(0.25*200Mbit/s+0.75*400Mbit/s)=175Mbit/sR2=0.5*(0.25*400Mbit/s+0.75*800Mbit/s)=350Mbit/sOpportunistic scheduling gives a 17% capacity gainBut what if both users need 200Mbit/s?
7 Cross-Layer Design TCP/IP-layers: Application Transport (TCP, UDP) Internet (IP)Network access (MAC, LLC)PhysicalToday: Some adaptation between neighbouring layersTomorrow: Network stack that take advantage of the interdependencies between the layers
8 Cross-Layer Design Variations follow different time scales: SNR variations ~ microsecondsCongestion of packets ~ seconds?Cumulative user traffic ~ secondsGoldsmith: Because of the different timescales, adaptation between layers is reasonable only if problems cannot be fixed locally within a layer.
9 What Influences Scheduling? QoS requirementsDelayThroughputSCHEDULINGFast fadingSlow fadingInstantaneous channel conditions
10 Restrictions From Other Layers Physical layer:Adaptive coding and modulation (MQAM)TDMA or CDMATDMA: The channel should be constant within a time-slotHigher layers:Throughput requirementsDelay requirements
11 Channel Models Two-state Markov: GOOD/BAD Slow fading: log-normal distributionFast fading:Rice (LOS)Rayleigh (LOS blocked)Nakagami (general)The average SNR, , for the fast fading models is influenced by slow fadingSlow scheduling: parameters changes slowlyFast scheduling: parameters changes fast
12 FairnessChoosing the best user with regard to the channel can lead to starvation of some usersFair algorithms assign a guaranteed time or throughput to the users (Robin Hood)Fairness not so important in a fast fading environment
13 Algorithms That Only Consider The Channel Conditions Proportional fair algorithmMax SNR schedulingMax SNR scheduling with a thresholdOpportunistic beamforming
14 Proportional Fair Algorithm Proportional fair if:increasing the current throughput by x% for one user leads to a cumulative throughput decrease for the other users of more than x%Maximises the product of the throughputsThe user with the relatively best channel is chosenStarvation is avoidedUsed by Qualcomm/HDR (IS-856)
15 Proportional Fair Algorithm STEP 1:At time t choose the user with the highest Ri(t)/ Ci(t):STEP 2:Update average rate:
16 Max SNR Scheduling Proportional Fair algorithm with: large values of tcsame for all users: Ri(t) ~ i(t) and Ci(t) ~The user with the largest SNR is chosenAlso called greedy algorithmNot fair if is different for different users
17 Max SNR Scheduling With Threshold The user with the largest SNR above a threshold is chosen:Reduces the amount of feedback from the users. Channel state information is only fed back if SNR is above the threshold.
18 Opportunistic Beamforming Induce channel fluctuations in a slow fading environment:MS’s with multiple antennasAntennas are fed with random phase and amplitudeThe overall induced SNR of a user is fed back to the BSThe BS schedules proportionally fair according to the different SNR values
19 Algorithms That Consider QoS FUNDAMENT: Revenue-based algorithmThis algorithm maximizes the throughput with regard to QoS requirements:wi(t): weight assigned to a user to include the QoS requirements
20 Algorithms That Consider QoS Examples of QoS-requirements:Minimum delay requirementMinimum throughput requirement
21 M-LWDF Modified Largest Weighted Delay First i: constant for controlling delay distributionsWi(t): head-of-the-line packet delaysThis simple algorithm is throughput optimal!
22 M-LWDF: Throughput Guarantees Guarantees minimum throughput Ri:Ri: constant token arrival rate in bucket iWi(t): delay of the longest delay token in bucketi: constant for controlling time-scale on which throughput guarantees are provided
23 Lazy SchedulerA scheduler that trades off delay for energy
24 M-LWDF: Delay vs. Power?i: constant for controlling trade-off between power and delayWi(t): head-of-the-line packet delaysPi(t): power that has to be provided to user i
33 System Model Buffering on both uplink and downlink SCHEDULER BUFFERS USER 1USER 2BUFFERSUSER 3BUFFERS
34 Current ResearchHave investigated buffering between wired and wireless (Rayleigh) networks using optimal SNR scheduling.New expression for overflow probability when the rate into the memoryless buffer is constant.The corresponding expression has been found for a queue with Poisson distributed traffic from the wired network (M/G/1-queue)
35 What Will I Do Next?Investigate max SNR Scheduling with both power and rate adaptation for M-QAMInvestigate the effects of Dopplerspread, coherence time and avg fade durationSpectral efficiency and BER for a scheduling algorithm using a -thresholdMax SNR scheduling with delayed CSISNR scheduling with different
36 Future Research 1) Physical layer issues: Optimising adaptive coding/modulation to the scheduling algorithm in useScheduling for slow fading channels (fairness!)Evaluating channel information inaccuracyAnalysing interference/multicell issuesCombining MIMO with schedulingCombining OFDM with schedulingEnergy efficient scheduling (with Sébastien)
37 Future Research 2) QoS issues: Must look at algorithms that: Provide QoS differentiation between usersMaximise the number of users that can be supported with the desired QoSProvide minimum flow guarantees?Gives minimum buffer overflow
38 Future Research 3) Other issues: Ad Hoc networks Multi-hop networks MAC and ARQ protocolsCAC: Connection Admission Control