Be Beyond nd Jai ains ns Fai Fairne ness Inde ndex: : Set - - PowerPoint PPT Presentation

▶
be beyond nd jai ain s n s fai fairne ness inde ndex set
SMART_READER_LITE
LIVE PREVIEW

Be Beyond nd Jai ains ns Fai Fairne ness Inde ndex: : Set - - PowerPoint PPT Presentation

Dec 08, 2023 115 likes •1.14k views

Be Beyond nd Jai ains ns Fai Fairne ness Inde ndex: : Set Settin ing The e Bar ar For or the e Dep eploy loymen ent of of Con ongest estion ion Con ontrol rol Alg lgorit orithms Ranysha Wa Ware Matthew K. Mu Ma

slide-1
SLIDE 1

Be Beyond nd Jai ain’s n’s Fai Fairne ness Inde ndex: : Set Settin ing The e Bar ar For

  • r the

e Dep eploy loymen ent

  • f
  • f Con
  • ngest

estion ion Con

  • ntrol

rol Alg lgorit

  • rithms

1

Ranysha Wa Ware Carnegie Mellon University Ma Matthew K. Mu Mukerjee Nefeli Networks Ju Justi stine S e Sher erry Carnegie Mellon University Sr Srinivasan Se Seshan Carnegie Mellon University

slide-2
SLIDE 2

I I hav have e desi designed gned a a new new CCA: : ! Ho How do do we we sho show w ! is is reason

  • nable

le to to deploy in th the Inte ternet? t?

2

slide-3
SLIDE 3

We typically use fairnes fairness to show that ! is reasonably deployable alongside ", a legacy algorithm.

3

slide-4
SLIDE 4

But ev every eryone

  • ne falls

falls short hort of achieving fair outcomes.

4

slide-5
SLIDE 5

But ev every eryone

  • ne falls

falls short hort of achieving fair outcomes.

5

Cubic can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.”

slide-6
SLIDE 6

But ev every eryone

  • ne falls

falls short hort of achieving fair outcomes.

6

CUBIC can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” BBRv1 can be unfair to Cubic, but “we are looking at modeling shallow buffer situations”.

slide-7
SLIDE 7

But ev every eryone

  • ne falls

falls short hort of achieving fair outcomes.

7

CUBIC can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” BBRv1 can be unfair to Cubic, but “we are looking at modeling shallow buffer situations”. PCC Vivace can be unfair to Cubic, but “as the number of CUBIC senders increases, it achieves the best fairness among new generation protocols.”

slide-8
SLIDE 8

But ev every eryone

  • ne falls

falls short hort of achieving fair outcomes.

8

CUBIC can be unfair to Reno, but “outside of TCP-friendly region” and “this doesn’t highly impact Reno’s performance.” Copa can be unfair to Cubic, but “is much fairer than BBR and PCC” and “uses bandwidth Cubic does not utilize.” BBRv1 can be unfair to Cubic, but “we are looking at modeling shallow buffer situations”. PCC Vivace can be unfair to Cubic, but “as the number of CUBIC senders increases, it achieves the best fairness among new generation protocols.”

slide-9
SLIDE 9

Ev Everyone makes excu excuses ses wh why th thei eir r algori rith thm is s sti still rea reason sonable le to to dep eploy loy desp espite ite unf unfair out utcomes.

9

slide-10
SLIDE 10

Th This is talk lk: We We need need a a pr pract actical cal depl deployment ent th threshold: a a bo bound und on n ho how w aggr aggressi essive e !, , a a new new CCA, , can can be be to ", , the he st stat atus us quo quo.

10 10

slide-11
SLIDE 11

Ou Outline: 1.

  • 1. What are desir

irable le prop

  • pertie

ies

  • f
  • f a deploy

loyment threshold

  • ld?

2.

  • 2. We defin

ine a new deploy loyment th threshold: harm.

11 11

slide-12
SLIDE 12

Ou Outline: 1.

  • 1. What are desir

irable le prop

  • pertie

ies

  • f
  • f a deploy

loyment threshold

  • ld?

2.

  • 2. We defin

ine a new deploy loyment th threshold: harm.

12 12

slide-13
SLIDE 13

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

13 13

PR PRACT CTICA CAL DEM DEMAN AND- AW AWAR ARE STA STATU TUS-QU QUO BI BIAS ASED ED MU MULTI TI- ME METR TRIC FU FUTURE- PR PROOF

slide-14
SLIDE 14

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

14 14

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-15
SLIDE 15

A deployment threshold needs to be pract practical: ical: should be feasible for new CCA to meet threshold.

15 15

slide-16
SLIDE 16

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

16 16

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-17
SLIDE 17

17 17

Slow bottleneck link

slide-18
SLIDE 18

18 18

CCA CCA: ! Slow bottleneck link

slide-19
SLIDE 19

19 19

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps La Latenc ncy: 5 ms CCA CCA: !

slide-20
SLIDE 20

20 20

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps La Latenc ncy: 5 ms CCA CCA: ! CCA CCA: "

slide-21
SLIDE 21

21 21

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps Do Downl nload d spe peed: d: 5 Mbps La Latenc ncy: 5 ms CCA CCA: ! CCA CCA: "

slide-22
SLIDE 22

22 22

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps Do Downl nload d spe peed: d: 5 Mbps La Latenc ncy: 5 ms 100 100 ms CCA CCA: ! CCA CCA: "

slide-23
SLIDE 23

23 23

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps Do Downl nload d spe peed: d: 5 Mbps

A deployment threshold needs to be mu multi lti-me metr tric ic: can account for performance metrics beyond just throughput.

CCA CCA: ! CCA CCA: " La Latenc ncy: 5 ms 100 100 ms

slide-24
SLIDE 24

24 24

Me Metrics l like l latency c y cannot b be “d “divided fa fairl rly”. ”.

slide-25
SLIDE 25

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

25 25

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-26
SLIDE 26

26 26

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps CCA CCA: !

slide-27
SLIDE 27

27 27

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps CCA CCA: !

slide-28
SLIDE 28

28 28

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps CCA CCA: ! CCA CCA: "

slide-29
SLIDE 29

29 29

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps 9 Mbps CCA CCA: ! CCA CCA: " Do Downl nload d spe peed: d: 1 1 Mb Mbps

slide-30
SLIDE 30

A deployment threshold needs to be st status-quo quo bias biased ed: based only

  • n impact of ! on ", not vice-versa.

30 30

Do Downl nload d spe peed: d: 1 1 Mb Mbps CCA CCA: ! CCA CCA: " Do Downl nload d spe peed: d: 10 Mbps 9 Mbps Li Link nk capa pacity: 10 Mbps

slide-31
SLIDE 31

31 31

Ja Jain’s f fairness i index i x is n not s status- quo quo bi biased. ased.

slide-32
SLIDE 32

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

32 32

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-33
SLIDE 33

33 33

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA CCA: !

slide-34
SLIDE 34

34 34

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA CCA: !

slide-35
SLIDE 35

35 35

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA CCA: ! CCA CCA: "

slide-36
SLIDE 36

36 36

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps Do Downl nload d spe peed: d: 7 Mbps CCA CCA: ! CCA CCA: "

slide-37
SLIDE 37

A deployment threshold needs to be dem demand and-aware aware: : do not penalize ! when " has inherently poor performance.

37 37

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 3 Mbps Do Downl nload d spe peed: d: 7 Mbps CCA CCA: ! CCA CCA: "

slide-38
SLIDE 38

38 38

Ma Max-mi min f fairness i is d dema mand a aware, equal equal-ra rate fa fairn rness is not.

slide-39
SLIDE 39

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

39 39

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-40
SLIDE 40

40 40

A deployment threshold needs to be future future-proof proof: useful on a future Internet where none of today’s current CCAs are deployed.

slide-41
SLIDE 41

41 41

A deployment threshold needs to be future future-proof proof: useful on a future Internet where none of today’s current CCAs are deployed.

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 1 Mbps CCA CCA: !

slide-42
SLIDE 42

42 42

A deployment threshold needs to be future future-proof proof: useful on a future Internet where none of today’s current CCAs are deployed.

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA CCA: !

slide-43
SLIDE 43

43 43

Does ! need to be nice to " and # or just "?

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA CCA: ! CCA CCA: "

slide-44
SLIDE 44

44 44

A future-proof threshold would only require ! to be nice to "

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps CCA CCA: ! CCA CCA: "

slide-45
SLIDE 45

45 45

TC TCP-fri friendliness is not fu future re-pr proof.

slide-46
SLIDE 46

We identify 5 des 5 desirable properties irable properties for a deployment threshold.

46 46

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-47
SLIDE 47

Ou Outline: 1.

  • 1. What are desir

irable le prop

  • pertie

ies

  • f
  • f a deploy

loyment threshold

  • ld?

2.

  • 2. We defin

ine a new deploy loyment th threshold: harm.

47 47

slide-48
SLIDE 48

When showing deployability: we run experiments of ! vs. " and me measure p perfor

  • rma

mance.

48 48

slide-49
SLIDE 49

49 49

throughput

When showing deployability: we run experiments of ! vs. " and me measure p perfor

  • rma

mance. An example

slide-50
SLIDE 50

50 50

throughput

When showing deployability: we run experiments of ! vs. " and me measure p perfor

  • rma

mance. Fairness compares these two bars

slide-51
SLIDE 51

51 51

throughput

When showing deployability: we run experiments of ! vs. " and me measure p perfor

  • rma

mance. Do not care what happens to "

slide-52
SLIDE 52

52 52

throughput

When showing deployability: we run experiments of ! vs. " and me measure p perfor

  • rma

mance. Only care about ! performance

slide-53
SLIDE 53

53 53

throughput

!

Only care about how ! performance changes

! alo alone

We want to me measure th the imp impact of " on ! performance.

slide-54
SLIDE 54

Ou Our Pr Proposal: De Deployment thre reshold should be based based on n ho how w much uch har harm ! do does es to to "

54 54

slide-55
SLIDE 55

55 55

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 10 Mbps La Latenc ncy: 5 ms CCA CCA: !

This is ! performance alone.

! alo alone

slide-56
SLIDE 56

56 56

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps Do Downl nload d spe peed: d: 5 5 Mb Mbps La Latenc ncy: 100 100 ms CCA CCA: ! CCA CCA: "

Ha Harm measures the impact of ! on " performance.

slide-57
SLIDE 57

57 57

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

slide-58
SLIDE 58

58 58

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!) !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand)

slide-59
SLIDE 59

59 59

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!)

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ": (")

! !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand) " = ! performance competing with "

slide-60
SLIDE 60

60 60

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!)

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ": (")

! !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand) " = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency):

! − " ! " − ! "

slide-61
SLIDE 61

61 61

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!)

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ": (")

! !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand) " = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency):

! − " ! " − ! "

$%%&' $%%

= .95

$%&' $%

= .50 Ex Example: : " caused throughput harm: " caused latency harm:

slide-62
SLIDE 62

62 62

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

Desirable threshold properties: ☐Practical ☐Demand-Aware ☐Status-Quo Biased Mu Multi-me metric ☐Future-Proof

Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!)

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ": (")

! !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand) " = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency):

! − " ! " − ! "

$%%&' $%%

= .95

$%&' $%

= .50 Ex Example: : " caused throughput harm: " caused latency harm:

slide-63
SLIDE 63

63 63

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

Desirable threshold properties: ☐Practical ☐Demand-Aware St Status-Qu Quo B Biased Multi-metric ☐Future-Proof

Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!)

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ": (")

! !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand) " = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency):

! − " ! " − ! "

$%%&' $%%

= .95

$%&' $%

= .50 Ex Example: : " caused throughput harm: " caused latency harm:

slide-64
SLIDE 64

64 64 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone: (!)

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ": (")

Harm is [0,1] where 0 is harmless and 1 is maximally harmful.

! !

Ho How t w to C Compute Ha Harm:

! = ! solo performance (demand) " = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm:

! − " ! " − ! "

$%%&' $%%

= .95

$%&' $%

= .50

Desirable threshold properties: ☐Practical De Demand-Aw Aware Status-Quo Biased Multi-metric ☐Future-Proof

slide-65
SLIDE 65

But But ho how w much uch har harm is s OK?

65 65

slide-66
SLIDE 66

Ke Key Insight: t: A A harm-based based thr hresho eshold: d: ! sho shoul uld d no not har harm " mu much mo more t than " har harms s itsel self

66 66

slide-67
SLIDE 67

67 67

Ha Harm(! vs

  • vs. ")
slide-68
SLIDE 68

68 68

Ha Harm(! vs

  • vs. ")

Ha Harm(" vs

  • vs. ")

?

slide-69
SLIDE 69

There are many possible thresholds based on harm (see paper!). One possible harm-based threshold: equiv equivalent alent-bounded bounded harm harm.

69 69

Ha Harm(! vs

  • vs. ")

Ha Harm(" vs

  • vs. ")

=

slide-70
SLIDE 70

One possible harm-based threshold: equiv equivalent alent-bounded bounded harm harm.

70 70

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps Do Downl nload d spe peed: d: 5 5 Mb Mbps La Latenc ncy: 100 100 ms CCA CCA: ! CCA CCA: "

Ha Harm(! vs

  • vs. ")
slide-71
SLIDE 71

71 71

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 5 Mbps Do Downl nload d spe peed: d: 5 5 Mb Mbps La Latenc ncy: 10 10 ms CCA CCA: ! CCA CCA: !

One possible harm-based threshold: equiv equivalent alent-bounded bounded harm harm.

Ha Harm(! vs

  • vs. !)
slide-72
SLIDE 72

72 72 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone:

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ":

! !

Ho How t w to C Compute Ha Harm:

x = ! solo performance (demand) y = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm:

! − # ! # − ! #

$%%&' $%%

= .95

$%&' $%

= .50

slide-73
SLIDE 73

73 73 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! !

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 10 10 ms ms

! ! vs

  • vs. !:

Ho How t w to C Compute Ha Harm:

x = ! solo performance (demand) y = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm: ! caused throughput harm: ! caused latency harm:

!"#$ !"

= .50

!"#$ !"

= .50

% − ' % ' − % '

!""#$ !""

= .95

!"#$ !"

= .50

! al alone: ! vs

  • vs. ":
slide-74
SLIDE 74

74 74 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone:

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ":

! ! ! ! vs

  • vs. !:

Ho How t w to C Compute Ha Harm:

x = ! solo performance (demand) y = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm: ! caused throughput harm: ! caused latency harm:

!"#$ !"

= .50

!"#$ !"

= .50

% − ' % ' − % '

!""#$ !""

= .95

!"#$ !"

= .50 Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 10 10 ms ms

slide-75
SLIDE 75

75 75 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone:

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ":

! ! ! ! vs

  • vs. !:

Ho How t w to C Compute Ha Harm:

x = ! solo performance (demand) y = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm: ! caused throughput harm: ! caused latency harm:

!"#$ !"

= .50

% − ' % ' − % '

!""#$ !""

= .95

!"#$ !"

= .50

!"#$ !"

= .50 Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 10 10 ms ms

slide-76
SLIDE 76

76 76 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone:

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ":

! ! ! ! vs

  • vs. !:

Ho How t w to C Compute Ha Harm:

x = ! solo performance (demand) y = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm: ! caused throughput harm: ! caused latency harm:

!"#$ !"

= .50

% − ' % ' − % '

!""#$ !""

= .95

!"#$ !"

= .50

Desirable threshold properties: Pr Practical Demand-Aware Status-Quo Biased Multi-metric ☐Future-Proof

!"#$ !"

= .50 Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 10 10 ms ms

slide-77
SLIDE 77

77 77 Do Download speed: : 10 Mbps La Late tenc ncy: 5 ms

! al alone:

Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 100 100 ms ms

! vs

  • vs. ":

! ! ! ! vs

  • vs. !:

Ho How t w to C Compute Ha Harm:

x = ! solo performance (demand) y = ! performance competing with " For “more is better” metrics (throughput): For “less is better” metrics (latency): Ex Example: : " caused throughput harm: " caused latency harm: ! caused throughput harm: ! caused latency harm:

!"#$ !"

= .50

% − ' % ' − % '

!""#$ !""

= .95

!"#$ !"

= .50

Desirable threshold properties: Practical Demand-Aware Status-Quo Biased Multi-metric Fu Future re-Pr Proof

!"#$ !"

= .50 Do Download speed: : 5 5 Mb Mbps La Late tenc ncy: y: 10 10 ms ms

slide-78
SLIDE 78

Is equivalent-bounded harm the answer? It It meet eets all all of

  • f our
  • ur crit

criteria. eria.

78 78

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF Fa Fairn rness and TCP CP-fri friendliness do not. DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-79
SLIDE 79

Is equivalent-bounded harm the answer? Bu But h t has i issues.

79 79

PR PRACT CTICA CAL STA STATU TUS-QU QUO BI BIAS ASED ED FU FUTURE- PR PROOF Fa Fairn rness and TCP CP-fri friendliness do not. DEM DEMAN AND- AW AWAR ARE MU MULTI TI- ME METR TRIC

slide-80
SLIDE 80

80 80

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 7 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA CCA: ! CCA CCA: !

slide-81
SLIDE 81

Could ! improve this imbalance? Equi

Equivalent nt-bo bounde unded d ha harm sa says n s no.

  • .

81 81

Li Link nk capa pacity: 10 Mbps Do Downl nload d spe peed: d: 7 Mbps Do Downl nload d spe peed: d: 3 Mbps CCA CCA: ! CCA CCA: "

slide-82
SLIDE 82

Other open questions:

  • 1. Alternatives to equivalent-bounded harm?
  • 2. Given a distribution of results, is there some ‘leeway in harm’? Should worry

about average or worst case results?

  • 3. What are the right workloads and networks for deployability testing?
  • 4. How widely deployed must a legacy CCA be in order to merit protection by our

threshold?

  • 5. If we have a threshold, should it be enforced? If so, how?

82 82

slide-83
SLIDE 83

Whi While e we we hav haven’t en’t set settled ed (yet et) on n th the perfect t th threshold, here is what what we we do do bel believ eve… e…

83 83

slide-84
SLIDE 84

Fa Fairn rness is is not

  • t wo

worki king ng as as a a pr pract actical cal thr hresho eshold. d.

84 84

slide-85
SLIDE 85

We We need need to st stop p maki aking ng ex excuses cuses fo for r why our r new algori rithms are re no not meet eeting ng an an unr unreal ealist stic c go goal al.

85 85

slide-86
SLIDE 86

Re Reasoning about harm rm is the ri right way fo forw rward rd to deri rive a new th threshold.

86 86

slide-87
SLIDE 87

87 87

Ranysha Wa Ware rware@cs.cmu.edu @ranyshware

Th The Bar ar For

  • r Deploy
  • yment: Do no more harm to the status quo

than it does to itself. Som Some op

  • pen question
  • ns:
  • 1. Alternative to equivalent-bounded harm?
  • 2. Given a distribution of results, is there some ‘leeway in harm’?

Should worry about average or worst case results?

  • 3. What are the right workloads and networks for deployability

testing?

Be Beyond nd Jai ain’s n’s Fai Fairne ness Inde ndex: : Set Settin ing The e Bar ar For

  • r the

e Dep eploy loymen ent

  • f
  • f Con
  • ngest

estion ion Con

  • ntrol

rol Alg lgorit

  • rithms
slide-88
SLIDE 88

BA BACKUP UP SL SLIDES

88 88

slide-89
SLIDE 89

Ev Every al algo gorithm hm is s unf unfai air?

89 89

slide-90
SLIDE 90

Example of unfair outcomes: Cubic is unfair to Reno.

90 90

slide-91
SLIDE 91

Example of unfair outcomes: Cubic is unfair to Reno.

91 91

slide-92
SLIDE 92

Example of unfair to outcomes: Cubic is unfair to Reno.

92 92

slide-93
SLIDE 93

What What is s TCP-fri friendliness?

94 94

slide-94
SLIDE 94

A mimicry-based threshold: If ! mimic mimics th the b behavior vior of " then ! is deployable.

TC TCP-fr friendliness: A TCP friendly flow should react to loss the same way that TCP Reno does such that

95 95

slide-95
SLIDE 95

TC TCP-fr friendliness: A TCP friendly flow should react to loss the same way that TCP Reno does such that

96 96

slide-96
SLIDE 96

What What do do you u mean ean by by st stat atus us-quo quo?

97 97

slide-97
SLIDE 97

There are some applications that are more popular than others.

98 98

Fi Figure: Internet Vi Video is already more than half of all Internet tr traffic

slide-98
SLIDE 98

Throughout this talk, this is how we defined harm:

99 99

Ha Harm(! vs

  • vs. ")

Ha Harm(" vs

  • vs. ")

?

slide-99
SLIDE 99

! "

In the paper, we define harm also as a function of the net network work condit conditions ions ! and and work workload load ".

100 100

Li Link nk capa pacity: 10 Mbps CCA CCA: # CCA CCA: $

slide-100
SLIDE 100

Ha Harm(! vs

  • vs. ",

, #, , $) Ha Harm(" vs

  • vs. ",

, #,

, $)

In the paper, we define harm also as a function of the net network work condit conditions ions # and and work workload load $.

101 101

?