1 of 25

Practical Fairness Metric for Network Resource Allocation

Presenter: Shaoting Feng

Mentor: Liangcheng Yu, Vincent Liu

2 of 25

Fairness matters

  • Fairness and performance isolation are important in many networking contexts

2

SJTU

3 of 25

Jain’s Fairness Index quantitatively measures fairness

  •  

3

SJTU

4 of 25

But, Jain’s Fairness Index is not enough!

  • JFI assumes ideal settings (e.g., stable, infinite-demand, synchronous connections)

​

​

4

SJTU

5 of 25

What are the practical applications?

  • Practical evaluation setting:
    • Asynchronous
    • Demand alteration
    • Finite demand
    • Changing network conditions
    • …

​

​

5

SJTU

6 of 25

6

02- Normalized Area Index

01- Extended JFI

03- Current project status

Outline

目录

CONTENTS

目 录

7 of 25

What if we divide the time?

  • Aggregate the JFIs and weight them based on the total throughput within the sub-region

7

SJTU

8 of 25

Have we solved the problem?

  • What does “fair” mean?

8

eJFI not fair

JFI fair

SJTU

9 of 25

Have we solved the problem?

  • What does “fair” mean?
  • Subsequent effects

9

SJTU

10 of 25

Have we solved the problem?

  • What does “fair” mean?
  • Subsequent effects

10

Dividing the sub-regions using breakpoints does not make sense

SJTU

11 of 25

What if we push to finest granularity?

  • The event of an application sending and receiving a single packet
  • Assuming information of application data input (output) events is pretty common
  • Think of the network as a whole

​

11

SJTU

12 of 25

12

02- Normalized Area Index

01- Extended JFI

03- Current project status

Outline

目录

CONTENTS

目 录

13 of 25

Design Goal

  • Boundedness
  • Scale and metric independence
  • Practical
  • Unbiased
  • Resilient to adversarial patterns

13

SJTU

14 of 25

Workflow

14

SJTU

15 of 25

Normalized Area Index

15

 

SJTU

16 of 25

Example use case: mul flows + async arrival

  • Configuration
    • 100 long-lived flows
    • 100 data bursts in separate periods
  • Variant: congestion control algorithm
    • CCA: TcpNewReno; TcpCubic

​

16

…

…

SJTU

17 of 25

Long/short flows reveal difference between CCA

17

Type

Tag

CCA

Demand

Period (s)

Long-lived flow

0-99

Experiment 1:

TcpNewReno

​

Experiment 2:

TcpCubic

2Mbps

0-10

Data burst

100-199

100Mbps

0-0.1

200-299

100Mbps

2-2.1

300-399

100Mbps

4-4.1

400-499

100Mbps

6-6.1

500-599

100Mbps

8-8.1

TcpNewReno: Unfairness Index = 0.10

TcpCubic: Unfairness Index = 0.41

SJTU

18 of 25

Long/short flows reveal difference between CCA

18

Type

Tag

CCA

Demand

Period (s)

Long-lived flow

0-99

Experiment 1:

TcpNewReno

​

Experiment 2:

TcpCubic

2Mbps

0-10

Data burst

100-199

100Mbps

0-0.1

200-299

100Mbps

2-2.1

300-399

100Mbps

4-4.1

400-499

100Mbps

6-6.1

500-599

100Mbps

8-8.1

TcpNewReno: Unfairness Index = 0.10

TcpCubic: Unfairness Index = 0.41

Our metric is capable of analyzing async flows with different life spans

SJTU

19 of 25

Example use case: change of data burst period

19

SJTU

20 of 25

As period increases, the network should be more fair.

  • Theory: The congestion window size of TcpCubic is irrelevant of RTT

​

​

​

​

​

  • Longer data bursts should be more fair.

20

SJTU

21 of 25

How to implement JFI?

21

SJTU

22 of 25

JFI doesn’t show the pattern; ours does.

22

SJTU

23 of 25

23

02- Normalized Area Index

01- Extended JFI

03- Current project status

Outline

目录

CONTENTS

目 录

24 of 25

Key of our metric: dynamic

  • Dynamic cases：
    • Long flows + short flows (an example of asynchronous)
    • Loss rate of channel
    • Node failures
    • Path changes
    • Demand changes
  • To identify fairness cases where
    • intuition and reality doesn’t match
    • theory can be deployed into real-world scenarios

​

24

SJTU

25 of 25

  • Normalized Area Index is a more practical metric to measure fairness
  • We need to deal with biased and gaming cases
  • We are working on interesting cases and applications to promote our work.

​

Thank you for your attention!

Summary

SJTU