Incentive-Oriented Downlink Scheduling forWireless Networks with Real-Time and

Non-Real-Time Flows

I-Hong Hou, Jing Zhu, and Rath Vannithamby

Motivation

• Wireless networks are increasingly used to serve real-time flows– VoIP, video streaming, online gaming

• In addition to throughput, these flows require strict per-packet delay guarantees

• Most current mechanisms belong to the paradigm of DiffServ

DiffServ

• Serve different flows differently• Usually, real-time flows get higher priorities

than non-real-time ones• Can be unfair to non-real-time flows• Non-real-time flows may lie about its category

to gain more service• Solution: charge real-time flows more

I am a real-time flow. I need small delay

You need to pay more

I require small throughput. Can I sacrifice throughput

for delay?

Goal of the paper

Design a scheduling policy that allows flows to tradeoff between high throughput and low delay by themselves

Desired Properties

• Incentive-compatibility: Flows optimize their own performance by reporting true category

• Versatility: The policy can work with various protocols in other layers– Different traffic patterns, different MAC, etc.

• Deadline awareness: The policy respects the deadlines of real-time flows

• Work conservation

Incentives of clients

• Non-real-time clients: Aim to maximize throughputs

• Real-time clients: Aim to maximize timely-throughput

• Timely-throughput: throughput of packets with delay < D

Basic Idea of Design

• Assume each client n has a weight of wn• Each client n is entitled to have wn/Σwn

channel time• Allocating channel time proportional to wn

maximizes and achieves proportional fairness• Deficit of client n: (The amount of channel

time that it is entitled) – (actual channel time)

Joint Deficit-Deadline (JDD) Policy

• A thin layer between Network layer and MAC layer

• Provides two functions: enqueue and dequeue– Interface defined by ns-2

• enqueue: a packet is labeled with deadline and put in the queue when it arrives from Network layer

• dequeue: forward a packet to MAC

Architecture of JDD

Enqueue

• When a packet arrives from Network layer• Mark the deadline of the packet

– Deadline of real-time flow = current time + D– Deadline of non-real-time flow = current time + a

large value (~ TCP timeout)

• Place the packet in the queue

Dequeue

• Delete all expired packets• Forward the packet with the earliest deadline

with the constraint that the corresponding client has positive deficit

• Earliest deadline: real-time packets usually got served first, and non-real-time packets need to wait

• Positive deficit: real-time packets receive smaller channel time

Ns-2 Simulation

• 10 clients, 5 of them are real-time ones, and the other 5 are non-real-time ones

• Real-time flows require a delay bound of 100ms

• All flows are generated by TCP• Use IEEE802.11 for the MAC• Weight of non-real-time clients = 1• Weight of real-time clients varies

Simulation Results

0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.90

0.51

1.52

2.53

3.54

Throughput of non-real-time

Timely-throughput of non-real-time

Timely-throughput of real-time

Weight of real-time client

Pe

rfo

rma

nc

e (

Mb

ps

)

Delay Distributions

0 0.1 0.2 0.3 0.4 0.5 0.60

0.2

0.4

0.6

0.8

1

1.2

Non-real-time

Delay (sec)

CD

F

Performance Comparison

• Compare against two other policies:– Earliest deadline first (EDF)– Weighted round robin (WRR)

• 10 clients, 5 real-time and 5 non-real-time• The distance between clients and the base

station is evenly distributed• Performance metric: • = throughput/timely-throughput

Simulation results for TCP

JDD WRR EDF0

500

1000

1500

2000

2500non-real-timereal-time

Pe

rfo

rma

nc

e (

kb

ps

)

JDD WRR EDF02468

101214161820

To

tal

we

igh

ted

lo

g

thro

ug

hp

ut

Simulation results for TFRC

JDD WRR EDF02468

101214161820

To

tal

we

igh

ted

lo

g

thro

ug

hp

ut

JDD WRR EDF0

500

1000

1500

2000

2500

3000

3500

non-real-time

real-time

Pe

rfo

rma

nc

e (

kb

ps

)

Simulation results for Interfering Network

• Add an additional link that causes interference

JDD WRR EDF02468

1012141618

To

tal

we

igh

ted

lo

g

thro

ug

hp

ut

JDD WRR EDF0

200400600800

10001200140016001800 non-real-time

real-time

Pe

rfo

rma

nc

e (

kb

ps

)

Conclusion

• We propose the JDD scheduling policy• The policy allows strategic clients to choose

between high throughput and low delay• The policy does not make any assumptions

on other layers, and can work with a wide range of different systems

• Simulation results show that JDD outperforms other policies