Stochastic Forecasts Achieve High Throughput and Low Delay ... · Stochastic Forecasts Achieve High Throughput and Low Delay over Cellular Networks Keith Winstein, Anirudh Sivaraman,

Stochastic Forecasts Achieve High Throughputand Low Delay over Cellular Networks

Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan

M.I.T. CSAIL

http://alfalfa.mit.edu

April 5, 2013

Cellular networks are variable

0

500

1000

1500

2000

0:00 0:30 1:00 1:30

thro

ugh

pu

t(k

bp

s)

Verizon LTE uplink throughput

Keith Winstein, Anirudh Sivaraman, and Hari Balakrishnan M.I.T. CSAIL http://alfalfa.mit.edu

Stochastic Forecasts Achieve High Throughput and Low Delay over Cellular Networks

Cellular networks are too reliable

0.2

0.3

2

3

0.1

1

10

0 50 100 150 200 250

Round-triptime(s)

Time (s)

(Verizon LTE, one TCP download.)



Interactive apps work poorly

I We measured cellular networks while driving:

I Verizon LTEI Verizon 3G (1xEV-DO)I AT&T LTEI T-Mobile 3G (UMTS)

I Then ran apps across replayed network trace:

I Skype (Windows 7)I Google Hangout (Chrome on Windows 7)I Apple Facetime (OS X)



Performance summary

0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)

Self-inflicted delay (ms)

Verizon LTE Downlink

Skype

Bette

r



Why is performance so bad?

I Exiting schemes react to congestion signals.

I Packet loss.I Increase in round-trip time.

I Feedback comes too late.

I The killer: self-inflicted queueing delay.

I Throughput overshoot means a queue filling up.



Sprout’s goal

I Most throughput

I Bounded risk of delay > 100 ms



Bounded risk of delay

I Infer link speed from interarrival distribution.

I Predict future link speed.

I Don’t wait for congestion.

I Control: Send as much as possible, but require:

I 95% chance all packets arrive within 100 ms.



Infer: link speed from flicker noise process

0.0001

0.001

0.01

0.1

1

10

100

(< 0.5)

39041 10 100 1000

Per

cen

tin

tera

rriv

als

interarrival time (ms)

t−3.27

(Verizon LTE, phone stationary.)



Predict: future link speed

I Model evolution of speed as random walk.

I (Brownian motion)

I Cautious forecast: 5th percentile cumulative packets

I Receiver makes forecast; sends back to sender in ack

I Almost all precalculated



Sprout’s model

Sender Queue Receiver

Poisson process

Rate λ controls

λ

σ

Brownian motion

λz

If in an outage,

drains queue

Poisson process

of σ√t varies λ λz is escape rate.



Parameters

Volatility σ: fixed @ 200 pkts/s√s

Expected outage time 1/λz : 1 sTick length: 20 msForecast length: 160 msDelay target: 100 msRisk tolerance: 5%

All source code was frozen before data collection began.



Control: fill up 100 ms forecast window

0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

Cautious fo

recast




0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

100 ms

Cautious fo

recast




0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

Sen

d 50

pkt

s

100 ms

Cautious fo

recast




0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

Sen

d 50

pkt

s

100 ms

100 ms

Cautious fo

recast




0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

Sen

d 50

pkt

s

Sen

d 5

100 ms

100 ms

Cautious fo

recast




0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

Sen

d 50

pkt

s

Sen

d 5

100 ms

100 ms100 ms

Cautious fo

recast




0

10

20

30

40

50

60

70

0 20 40 60 80 100 120 140 160

Cum

ulat

ive

pack

ets

sent

time (ms)

Sen

d 50

pkt

s

Sen

d 5

Sen

d 3 100 ms

100 ms100 ms

Cautious fo

recast



0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)



Skype

Bette

r



0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)



SkypeFacetime

Google Hangout

Bette

r



0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)



Compound TCP

LEDBAT

SkypeFacetime

Google Hangout

Vegas

Bette

r



0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)



Compound TCP

LEDBAT

Cubic

SkypeFacetime

Google Hangout

Vegas

Bette

r



0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)



Compound TCP

LEDBAT

Cubic

SkypeFacetime

Google Hangout

Vegas

Sprout

Bette

r



0

1000

2000

3000

4000

5000

6000

7000

8000

100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)



Sprout-EWMA

Compound TCP

LEDBAT

Cubic

SkypeFacetime

Google Hangout

Vegas

Sprout

Bette

r



0

1000

2000

3000

4000

5000

300500100020003000500010000

Th

rou

gh

pu

t(k

bp

s)


Verizon LTE Uplink

Sprout-EWMA

Compound TCP

LEDBAT

Cubic

Skype Facetime

Google Hangout

Vegas

Sprout



150

200

250

300

350

400

450

500

550

200030005000100002000050000100000

Th

rou

gh

pu

t(k

bp

s)


Verizon 3G (1xEV-DO) Downlink

Sprout-EWMACompound TCP

LEDBAT

Cubic

Skype

FacetimeGoogle Hangout

Vegas

Sprout



200

300

400

500

600

200300500100020003000500010000

Th

rou

gh

pu

t(k

bp

s)


Verizon 3G (1xEV-DO) Uplink

Sprout-EWMA

Compound TCP

LEDBAT

Cubic

Skype

Facetime

Google Hangout

Vegas

Sprout



500

1000

1500

2000

2500

3000

3500

4000

3050100200300500100020005000

Th

rou

gh

pu

t(k

bp

s)


AT&T LTE Downlink

Sprout-EWMA

Compound TCPLEDBATCubic

Skype

Facetime

Google Hangout

Vegas

Sprout



200

300

400

500

600

700

800

900

2003005001000200050001000020000

Th

rou

gh

pu

t(k

bp

s)


AT&T LTE Uplink

Sprout-EWMA

Compound TCP

LEDBATCubic

SkypeFacetime

Google Hangout

Vegas

Sprout



200

400

600

800

1000

1200

1400

1600

3005001000200030005000

Th

rou

gh

pu

t(k

bp

s)


T-Mobile 3G (UMTS) Downlink

Sprout-EWMA

Compound TCPLEDBAT

Cubic

SkypeFacetime

Google Hangout

Vegas

Sprout



200

300

400

500

600

700

800

900

1000

2003005001000200050001000030000

Th

rou

gh

pu

t(k

bp

s)


T-Mobile 3G (UMTS) Uplink

Sprout-EWMA

Compound TCP

LEDBAT

Cubic

Skype

Facetime

Google Hangout

Vegas

Sprout



Overall results

Sprout vs. Avg. speedup Delay reduction

Skype 2.2× 7.9×Hangout 4.4× 7.2×Facetime 1.9× 8.7×Compound 1.3× 4.8×TCP Vegas 1.1× 2.1×LEDBAT Same 2.8×Cubic 0.91× 79×



Varying risk tolerance

200

300

400

500

600

700

800

900

1000

2003005001000300050001000030000

Th

rou

gh

pu

t(k

bp

s)


Compound TCP

LEDBAT

Cubic

Skype

Facetime

Google Hangout

Vegas

Sprout (5%)

25%

50%

75%

95%



Competes with AQM even though end-to-end

0

20

40

60

80

100

2003005001000300050001000030000

Uti

lizat

ion

(per

cen

t)


Sprout

Sprout-EWMA

Cubic

Cubic-over-CoDel



Competing traffic inside Sprout tunnel

Direct via Sprout Benefit

Cubic throughput 8336 kbps 3776 kbps 0.5× (= worse)Skype throughput 78 kbps 490 kbps 6×Skype 95% delay 6.0 s 0.17 s 35×



Replication by Stanford students (February–March 2013)

I Alterman & Quach reproduced some of our measurements

I http://ReproducingNetworkResearch.wordpress.com/2013/03/12/1216/

I Won best project award in Stanford networking class!



M.I.T. 6.829 contest (March–April 2013)

I Turnkey network emulator, evaluation

I Sender, receiver run in Linux containers

I 4th prize: $20

I 3rd prize: $30

I 2nd prize: $40

I (If beat Sprout) 1st prize:

Co-authorship on future paper



M.I.T. 6.829 contest (March–April 2013)

I Turnkey network emulator, evaluation

I Sender, receiver run in Linux containers

I 4th prize: $20

I 3rd prize: $30

I 2nd prize: $40

I (If beat Sprout) 1st prize: Co-authorship on future paper



Baseline



Land of 3,000 student protocols



Sprout is on the frontier



Limitations

I Only evaluated long-running flows.

I All testing data from Boston.

I User should wrap competing flows inside Sprout.

I If queue is full of another user’s packets, an end-to-endscheme can’t help.

I Fortunately, cells have per-device queues. . .I . . . but Wi-Fi generally doesn’t.

I What about when the cell link isn’t the bottleneck?



Our approach

I Pick a model, any model.

I All models are wrong, but they help anyway!

I See if it lands on the frontier.*

* (On a large set of real network paths or newly-collectedtraces.)

I Kaizen for congestion



Thank you

I Lakshminarayanan Subramanian

I Shuo Deng

I Jonathan Perry

I Katrina LaCurts

I Andrew McGregor

I Tim Shepard

I Dave Taht

I Michael Welzl

I Hannes Tschofenig

I Wireless@MIT members (http://wireless.csail.mit.edu)

I NSF & Shannon family (fellowship)



Sprout for controlled delay over cellular networks

I Infer link speed from interarrival distribution

I Predict future link speed

I Control risk of large delay with cautious forecast

I Yields 2–4× throughput of Skype, Facetime, Hangout

I Achieves 7–9× reduction in self-inflicted delay

I Matches active queue management without router changes

I Code and directions at http://alfalfa.mit.edu

I [email protected]



Documents

Stochastic Forecasts Achieve High Throughput and Low Delay ... · Stochastic Forecasts Achieve High Throughput and Low Delay over Cellular Networks Keith Winstein, Anirudh Sivaraman,