Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom...

37
Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC), Shao Liu (UIUC), Laurent Massoulié (MSRC) MIT, 09 Nov 05
  • date post

    19-Dec-2015
  • Category

    Documents

  • view

    215
  • download

    1

Transcript of Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom...

Page 1: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

Farsighted Congestion Controllers

Milan Vojnović

Microsoft Research

Cambridge, United KingdomCollaborators:Dinan Gunawardena (MSRC), Peter Key (MSRC), Shao Liu (UIUC), Laurent Massoulié (MSRC)

MIT, 09 Nov 05

Page 2: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

2

Problem

Applications concerned with long-run throughput Indifferent to short-timescale throughput Ex. peer-to-peer file sharing

Goal: Optimize long-run throughput Share bandwidth fairly with TCP

Data transfer

WebWeb Internet

Page 3: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

3

0 0.5 1 1.5 2

x 104

0

1

2

3

4

5

6

x 106

time

rates

rates over time, FAR and TCP

FAR

TCP

Solution

Number ofconnections

Farsighted TCP

TCPTCP

Rat

e (M

b/s)

Internet

Time

Page 4: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

4

Solution: farsighted controller

w w + 1/ww max(w – 1/(ww0)

+ ack- ack

-m

Window

Time

high congestion

• Two-timescale control• = parameter learned on-line at slow timescale

w0

Page 5: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

5

Compare with TCP

Window

Time

w w + 1/ww w – ½ w

+ ack- ack

high congestion

Page 6: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

6

Roadmap

Optimality Properties Rate adaptation Protocol & verification Conclusion

Page 7: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

7

Setup

Network state fluctuates over a set of phases U

Ex. single link phase = number of competing flows

(u) = fraction of time phase is u Cl,u(x) = cost of link l with arrival rate x

Network

Page 8: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

8

Setup (cont’d)

Vr(x) = utility for rate x = (x(u), uU)

User r

Uu

rrrr uxUuxV ))(()()(

rrr xxU / const )(

)()( rrrr xUxV

Uu

rr uxux )()(

TCP-like Long-run throughput optimizer

Page 9: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

9

Problem

l ql

qulUuRr

rr uxCuxV ))(()()( ,

0)(uxr

maximize

over

SYSTEM:

Rrxr , optimal if it solves SYSTEM

Page 10: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

10

TCP-like only

l ql

qulRr

rr uxCuxUu

))(())(( ,

0)(uxr

maximize

over

• Separation into independent problems

• Traditional controllers are “myopic”• Optimize rates “independently over time”

SYSTEM u:

Page 11: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

11

With long-run throughput optimizers

l qlqul

Uu

Frrr

Mrrr

uxCu

xUuxUu

))(()(

)())(()(

,

0)(uxr

maximize

over

• No separation

• Long-run throughput optimizers = “farsighted”

SYSTEM:

Page 12: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

12

Formally: multi-path problem

phase 1 phase 2 phase 3 phase N. . .

rxr(1) xr(2) xr(3) xr(N)

Studied by Gibbens & Kelly 02

But our setup in phase spacePath is not spatial path present at all times “Paths come and leave over time”Time (not space) diversity

Page 13: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

13

Roadmap

Optimality Properties Rate adaptation Protocol & verification Conclusion

Page 14: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

14

Price equalization

Farsighted user r pr(u) = price when phase is u (price = loss event rate)

rr

rrr

pup

pupux

)( else,

)( ,0)( If

)(' rrr xUp

“good phase”

“bad phase”

“reference price”

Page 15: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

15

Special: single link

farsightedmyopic 1

u

Phase u = u competing myopic flows

xF(u)xM(u)

else

)()(

0

1 uuuxuxF

else)(

u

uuxuxM 1

)),max()(()( '' u

FM uxuUxU 01

1 uxu : integer largest

capacity = 1

Page 16: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

16

Farsighted users are conservative

A flow said r to be conservative iff

= average user-perceived price

)(1'

rrr pUx

rp

ur

urr

r uxu

uxupup

)()(

)()()(

Seen as throughput maximizers under a “TCP-friendly” constraint

“TCP-friendly”

If TCP lossthroughput(C)

Farsighted user: “=“ in (C)

Page 17: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

17

Throughput comparison

Consider a farsighted user F and a myopic user M

Both with same utility functions Both competing for same set of links

MF xx Result

Page 18: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

18

Diminishing returns with switching to farsighted n flows k farsighted, n-k myopic flows use same routes = throughput of farsighted flow for given k

kkxF withdecreases )(

)(kxF

Result

Page 19: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

19

Can be made “low-priority”

One link characterized by increasing, convex function

Strictly concave utility functions f farsighted flows (0) = fraction of time no myopic flow on the link

Result

0 all ),()( ')0(' xxUxU MfF

“low-priority” iff

Page 20: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

20

File transfer time Short-lived flows:

Poisson arrivalsExponential file sizes

short lived

long lived myopic

S1:

short lived

long lived farsighted

S2:

21 TT Result Ti = mean file transfer time in Si

Page 21: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

21

Roadmap

Optimality Properties Rate adaptation Protocol & verification Conclusion

Page 22: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

22

Traditional myopic

))()(( ' dtNxdtxUxkdx rrrrrrr 212

))(( '

rl

lrrrrdtd qxUkx ql = price at link l

Fast time scale (RTT)

TCP:• 0 or 1• 1 with rate

rllr qx

const

rl

lq)('rr xU

Page 23: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

23

Farsighted

))(( 'rrrrrdt

d xUa

1

)(

rl

lrrrdtd qkx

Fast timescale (RTT)

Slow timescale

ar small r)('rr xU

Page 24: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

24

Roadmap

Optimality Properties Rate adaptation Protocol & verification Conclusion

Page 25: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

25

Back to the solution

w w + 1/ww w – 1/(w

+ ack- ack

-m

Window

Time

high congestion

• Two-timescale control• = parameter learned on-line at slow timescale

Page 26: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

26

Sensing phase

vcwnd vcwnd + 1/w0

vcwnd vcwnd – 1/(w0+ ack- ack

-m

Time

w0

Sequential hypothesis testing: p In fact, optimal for Poi(pw0) losses (CUSUM)

Know how to set m so false positives are rare and control is responsive (reflected random walk)

Page 27: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

27

“Reference price”: initial guess Want be almost constant Solution: small gain for adaptation But need to converge to equilibrium Solution:

Initial guess = current loss rate

gain

number of iterates

g_min

g_max

n0 n1

loss rate

g_max = 0.005g_min = 0.0001

Page 28: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

28

Verification by simulation Scenario 1:

1 2 3 4 5 6 7 8 9 100

1

2

3

4

5

6

7

phases

flow

num

ber

Pyrimid Topology, 2-6 flows

1 period has 9 phasesu = (2,3,4,5,6,5,4,3,2)

RED, 6 Mb/sLong-lived farsightedLong-lived TCP

Phase duration = 800 sec

Page 29: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

29

Send rate

0 1000 2000 3000 4000 5000 6000 7000

0

0.5

1

1.5

2

2.5

3

3.5

4

4.5

x 106 Rate vs time, FAR and TCP

time

rate

FAR

TCP

Time (sec)

Sen

d ra

te (

Mb/

s)

FAR

TCP

Page 30: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

30

Loss rate

0 1000 2000 3000 4000 5000 6000 70000

1

2

3

4

5

6

7

8

x 10-3 Loss event probabilities of FAR and TCP and xi of FAR

time

Pro

bability

FAR

TCPxi

Loss

rat

e

Time (sec)

FAR

TCPReference loss rate

Page 31: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

31

Per phase rate averages

2 2.5 3 3.5 4 4.5 5 5.5 6 6.5 70

0.5

1

1.5

2

2.5

3

3.5

4

4.5x 10

6Phase average rate: Theoretical Vs Measurement, Farsighted and TCP

Phases

Pha

se a

ve ra

tes

Far measured

Far theoreticalTCP measured

TCP theoretical

The 7th phase is theaverage value over allphases

Phase FAR (Mbps) TCP (Mbps)

2 4.38/4.24 1.61/1.73

3 2.77/2.46 1.61/1.77

4 1.15/1.20 1.61/1.58

5 0/0.62 1.50/1.33

6 0/0.23 1.20/1.11

Avg rate

1.61/1.73 1.53/1.53

Phase

FAR theory

FAR simulation

TCP simulations

TCP theory

Total Avg

Ave

rage

sen

d ra

te (

Mb/

s)

Page 32: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

32

Scenario 2

0.2 0.4 0.6 0.8 1 1.2 1.4 1.6 1.8 2 2.2

x 104

2

3

4

5

6

7

8

time

flow

num

ber

inclu

din

g p

ers

iste

nt

flow

s

flow number over time

0 0.5 1 1.5 2

x 104

0

1

2

3

4

5

6

x 106

time

rate

s

rates over time, FAR and TCP

FAR

TCP

1 2 3 4 5 6 7 80

1

2

3

4

5

6

7x 10

6 Phase average rate: Theoretical Vs Measurement, Far and TCP

phases

phas

e av

e ra

te

FAR measurement

FAR theoreticalTCP measurement

TCP theoretical

The last phase is for theaverage of all phases

Time (sec)

Num

ber

of F

low

sS

end

Rat

e (b

/s)

Ave

rage

sen

d ra

te (

Mb/

s)

Time (sec)Phase

FAR theory

FAR simulation

TCP simulations

TCP theory

Total Avg

Page 33: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

33

File transfer time

RED, 6 Mb/s

TCPTCP

RED, 6 Mb/s

FARTCP

Fn ~ Exp()

Tn = Poi()

S1:

S2:

= 0.11/ = 10 MB

S1 S2

Avg Flow Number 8.7139 8.1679

Avg file transfer time (sec) 179 173

Avg link bandwidth (Mb/s) 10.80 10.82

Per connection avg rate (Mb/s) TCP = 1.3405

TCP = 1.3472

FAR = 1.3642

TCP = 1.3262

Page 34: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

34

Benefits to other flows?

Ex. same as earlier slide But 10 long-lived flows: either all TCP or all FAR

= 0.051/ = 20 MB

10 FAR 10 TCP

Avg Flow Number 6.92 12.84

Avg Transfer Time (sec) 349 470

Page 35: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

35

More realistic traffic

Synthetic web (UNC, Jeffay+) Requests, responses,

idle times drawn from empirical distributions

S1: 1 persistent TCP S2: 1 persistent FAR

Both S1 & S2: number of web users = 1

0 100 200 300 400 500 6000

100

200

300

400

500

600

TCP

FAR

File transfer time for FAR and TCP

TCP: File transfer time (sec)

FA

R:

File

tra

nsfe

r tim

e (s

ec)

Page 36: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

36

Conclusion

Farsighted Congestion Control Solution for long-run throughput optimization

Decentralized control No special feedback required

(standard TCP sender modif) Not a heuristic hack

Microeconomics rationale Benefits to other flows On-going:

Further simulations Experimental implementation in MS Vista Real-word experiments

Page 37: Farsighted Congestion Controllers Milan Vojnović Microsoft Research Cambridge, United Kingdom Collaborators: Dinan Gunawardena (MSRC), Peter Key (MSRC),

37

More

http://research.microsoft.com/~milanv/farsighted.htm

& Thanks!