1

Distributed Fair Scheduling in a Wireless LAN

Nitin Vaidya, Texas A&M University

Victor Bahl, Microsoft Research

Seema Gupta, now with Cisco

MobiCom 2000

2

Distributed Scheduling :

What & Why

3

Medium Access Control

Wireless medium is a broadcast medium

Transmissions by multiple nodes can interfere

Need medium access control (MAC)

Many proposals Centralized Distributed

4

BaseStation

Node1

Node2

Node3

Noden

Centralized Protocols

Base station coordinates access to the wireless channel

5

Distributed Protocols

All nodes have identical responsibilities

Wireless LAN

Node1

Noden

Node3

Node2

6

Disadvantages of Centralized Approach

If a node cannot talk to the base station, it cannot transmit to any other nodes

Base station needs to keep track of state of other nodes

Hard to use failure-prone nodes as coordinators in centralized protocols

7

Fairness

8

Fairness

Packets to be transmitted belong to several flows

Each flow is assigned a weight

Bandwidth assigned to each backlogged flow is proportional to its weight

9

Fairness

Three flows with weights 2, 1, 1

Backloggedflows:

Allocatedbandwidth

10

Fair Queueing

Many centralized fair queueing protocols exist WFQ, WF2Q, SCFQ, SFQ, …

Scheduler needs to know state of all flows

Flow 1

Flow 2

Flow n

Output link

11

Distributed Fair Scheduling

12

Our Objectives

Fully distributed fair scheduling protocol All nodes have identical responsibilities

Nodes do not need to be aware of each other’s state

Maintain compatibility / resemblance with an existing standard specifically, IEEE 802.11 Distributed Coordination Function

13

Proposed Approach

Combination of

IEEE 802.11 Distributed Coordination Function (DCF) Carrier sense / collision avoidance

A centralized fair queueing protocol

14

Basic Carrier Sense Approach

A node wishing to transmit waits until channel is sensed as idle, and then transmits

If two nodes are waiting to transmit, they will collide

Collision avoidance mechanism needed to avoid this

15

IEEE 802.11 Distributed Coordination Function

Collision avoidance mechanism: When transmitting a packet, choose a backoff interval in the range [0,cw]

– cw is contention window

Count down the backoff interval when medium is idle

When backoff interval reaches 0, transmit

0 cw

16

802.11 DCF Example

data

waitB1 = 5

B2 = 15

B1 = 25

B2 = 20

data

wait

B1 and B2 are backoff intervalsat nodes 1 and 2cw = 31

B2 = 10

17

Self-Clocked Fair Queueing (SCFQ) [Golestani]

A centralized fair scheduling protocol

But more amenable for a distributed implementation than many others

The steps involved in deriving proposed distributed protocol starting from SCFQ are given in the paper virtual time, start/finish tags implementation does not need virtual time or tags

18

Distributed Fair Scheduling (DFS)

Node with smallest “length/weight” should transmit first Caveat: This is a somewhat imprecise statement. DFS

(implicitly) compares so-called virtual finish tags, which are a function of length/weight

See paper for details on the finish tags

Backoff intervals used as a way to distributedly determine whose “length/weight” is smaller

19

Distributed Fair Scheduling (DFS)

Choose backoff interval

= packet length / weight

packet length = 5

weight = 1/3

backoff interval = 5 / (1/3) = 15 slots

20

Distributed Fair Scheduling (DFS)

data

wait

B1 = 15

B2 = 5

Packet length = 15

Weight of node 1 = 1 ====> B1 = 15 / 1 = 15Weight of node 2 = 3 ====> B2 = 15 / 3 = 5

B1 = 10

B2 = 5

data

wait

B1 = 5

B2 = 5

Collision !

21

Collisions

Collisions occur when two nodes count down to 0 simultaneously In centralized fair queueing, ties can be broken without causing

“collisions”

To reduce the possibility of collisions:

Backoff interval = Scaling_Factor * length / weight * random number with

mean 1

22

Backoff Interval

Initial formula: Length / weight = 15 / 1 = 15

Scaling_factor * length / weight * random number

= 4 * 15 / 1 * [0.9,1.1]

= [54,66]

0 15

23

Backoff Interval

802.11

Proposed DFS

0

0

24

Collisions Resolution

Collision occurs when two nodes count down to 0 simultaneously

Counting to 0 implies that it is a given node’s “turn” to transmit

To reduce “priority” reversals, a small backoff interval is chosen after the first collision

Backoff interval increased exponentially on further collisions

25

Impact of Small Weights

Backoff interval: Scaling_factor * length / weight * random

number

Backoff intervals can become large when weights are small

Large backoff intervals may degrade performance (time wasted in counting down)

26

Impact of Small Weights

Recall: Backoff intervals are being used to compare “length/weight”

Intuition: Any non-decreasing function of lenghth/weight may be used to obtain backoff intervals

27

Alternative Mappings

Scaling_factor * length / weight * random number

Chosenbackoffinterval

Linear mapping

EXP

SQRT

28

Alternative Mappings

Advantage smaller backoff intervals less time wasted in counting down when weights of all

backlogged flows are small

Disadvantage backoff intervals that are different on a linear scale may

become identical on the compressed scale possibility for greater number of collisions

29

Performance Evaluation

Using modified ns-2 simulator: 2 Mbps channel

Number of nodes = N Number of flows = N/2 Odd-numbered nodes are destinations,

even-numbered nodes are sources

Unless otherwise specified: flow weight = 1 / number of flows backlogged flows with packet size 584 bytes (including UDP/IP headers)

Scaling_Factor = 0.02

30

Fairness Index

Fairness measured as a function of

(throughput T / weight ) for each flow f over an interval of time Unless specified, the interval is 6 seconds

31

Throughput / Weight Variation Across Flows (with 16 Flows)

Flow destination identifier

Throughput / Weight

Flattercurve

is fairer

DFSis fairer

802.11

32

Throughput - Fairness Trade-Off

Aggregatethroughput(all flowscombined)

Number of flows

802.11

33

Throughput - Fairness Trade-Off

Fairnessindex

Number of flows

802.11

34

Scaled 802.11

Fairness of 802.11 can be improved by using larger backoff intervals

Is DFS fairer simply because it uses large backoff intervals ?

Scaled 802.11 = 802.11 which uses backoff

interval range comparable with DFS

35

Short-Term Fairness

Number of packets transmitted by a flow (over 0.04 second windows)

Frequency

Narrowdistribution

is fairer

DFS isfairer

DFS

Scaled 802.11

802.11

36

Fairness versus Sampling Interval Size(24 flows)

Interval Size

Fairnessindex

DFS

Scaled 802.11

802.11

37

Alternative Mappings for Backoff Intervals

See additional data in the paper

EXP and SQRT improve throughput compared to LINEAR mapping when all backlogged flows have low weights but not too impressively

If at least one backlogged flow has a high weight, not much benefit

38

Conclusions(supporting arguments for some conclusions not

presented in the talk: please see the paper)

DFS improves fairness compared to 802.11 and Scaled 802.11

Alternative mappings somewhat beneficial

No distributed fair scheduling protocol may accurately emulate work-conserving centralized protocols (unless clocks are synchronized)

39

Conclusions

Possible to augment DFS with other techniques to improve fairness in presence of transmission errors see Seema Gupta’s M.S. thesis

No performance cost even if weight assigned to a flow is changed on a per-packet basis Execution complexity of centralized protocols would

increase

Possible to handle multiple flows per node

40

Other Potential Applications of DFS

Wired LANs

Wireless multi-hop networks see our 1999 Microsoft Research technical report for some

initial ideas

41

Issues for Further Work

DFS is only the first step towards practical fairness:

How to choose parameters such as Scaling_Factor ? Failure to choose reasonable values can degrade throughput or short-

term fairness

How to choose flow weights ? Let upper layer specify dynamically, or Static assignment based on static criteria

Ad hoc network-related issues

42

Thank you!

www.cs.tamu.edu/faculty/vaidya

43

Thank you!

www.cs.tamu.edu/faculty/vaidya

44

Impact of Packet Size

Three flowswith differentpacket sizes

Packet size (bytes)

584 328 200

Flowthroughput

802.11

45

Impact of Scaling Factor(six flows with weights 1/2,1/4,1/8,1/16,1/32,1/32)

Fairnessindex

Scaling Factor

DFS

46

Impact of Scaling Factor (six flows with weights 1/2,1/4,1/8,1/16,1/32,1/32)

Scaling factor

Aggregatethroughput

DFS

47

Related Past Work

Centralized fair queueing on wired links [Bennett,Demers,Parekh]

Centralized fair queueing in wireless environments, taking location-dependent errors into account [Bharghavan,Ramanathan,Zhang]

Distributed Real-time scheduling [Sobrinho] Distributed Priority-based scheduling

48

Backoff Interval

Scaling factor Small number : May result in more collisions Large number: Larger overhead

Random number range Small range will cause more collisions between synchronized

nodes

How to choose these adaptively ? This paper punts the issue But heuristic solutions are easy to define Heuristics yet to be evaluated