characterizing the transport behaviour of the short

Characterizing the Transport Behaviour of the

Short Message ServiceEarl Oliver

David R. Cheriton School of Computer ScienceUniversity of Waterloo

MobiSys 2010June 17, 2010

1

Tuesday, September 14, 2010

Takeaways

2


Takeaways

• Studied the channel characteristics of SMS from the perspective of mobile devices sending bursts of messages (bulk senders)

2


Takeaways

• Studied the channel characteristics of SMS from the perspective of mobile devices sending bursts of messages (bulk senders)

• Design and implement an efficient and reliable SMS-based data transport protocol

2


SMS in mobile systems

• Prolific use of SMS

3




• ubiquity

3




• ubiquity

• reliability

3




• ubiquity

• reliability

• low-cost*

3




• ubiquity

• reliability

• low-cost*

• convenient APIs

3




• ubiquity

• reliability

• low-cost*

• convenient APIs

• endpoint addressability

3


Existing use of SMS

• Typically constrained to single messages

4


Existing use of SMS


• Stop-and-wait protocol used to send larger amounts of data

4


Existing use of SMS



• Easy to implement

4


Existing use of SMS



• Easy to implement

• Network communication often a secondary concern

4


Alternatives to SMS

5


Alternatives to SMS

• Cellular data services

5


Alternatives to SMS


• low-latency, high data rate

5


Alternatives to SMS



• end-points behind NAT

5


Alternatives to SMS




• sparsely deployed in developing regions

5


Alternatives to SMS





• MMS, EMS

5


Alternatives to SMS





• MMS, EMS

• can transfer large amounts of data

5


Alternatives to SMS





• MMS, EMS

• can transfer large amounts of data

• poorly supported and not universally available

5


Goal

• Develop a portable library to exchange large mounts of data over SMS

6


Goal


• How do we efficiently maximize the use of the SMS channel?

6


Goal



• Minimize message overhead

6


Goal




• Maximize data throughput

6


Outline

• Channel characterization

• Transport protocol design

• Evaluation

7


Source Destination

Service center

8

Previous work

• Zerfos (IMC 2006)


Source Destination

Delay and loss rate

Service center

8

Previous work

• Zerfos (IMC 2006)


Source Destination

Service center

Our work

• We examine channel properties from the perspective of mobile devices using the service as mass message senders

9


Source Destination

Transmission time, delay, loss rate,

message reordering

Transmission time, delay, loss rate,

message reordering

Black box

Our work

• We examine channel properties from the perspective of mobile devices using the service as mass message senders

9


Sources of variability

Source Destination

Base station andcontroller


Service center

10



Source Destination



Service center

Dedicatedchannel request

10



Source Destination



Service center

Dedicatedchannel request

ALOHA

10



Source Destination



Service center

Channel allocation

10



Source Destination



Service center

Message transmit

10



Source Destination



Service center

10



Source Destination



Service center

Transmit time

10



Source Destination



Service center

Transmit time

Page and dedicatedchannel request

10



Source Destination



Service center

Transmit time

Channel allocation

10



Source Destination



Service center

Transmit time

10



Source Destination



Service center

Transmit timeDelay

10



Source Destination



Service center

Transmit timeDelay

Loss rate,reordering rate

10


Experiment

11

Transmission time

Delay

Loss rate

Reordering rate

Intra-burst time

Transmission index

Network interface

Time-of-day} }{{ X

Channel characteristics Experimental parameters


Experiment

1234

Intra-burst time

Transmission index

Network interface

Time-of-day

12


Methodology

13


Methodology

• Transfer repeated bursts of messages between pairs of stationary BlackBerrys and USB tethered Nokia cell phones

13


Methodology


• Major Canadian GSM provider (Rogers)

13


Methodology



• 80,000 messages

13


Methodology



• 80,000 messages

• See paper for details of variable isolation and experiment methodology

13


Summary of key resultsFinding Design impact

Transmission timeTransmission timeTransmission time is

independent of intra-burst time, index, time-of-day.

Protocol does not need to regulate message

transmission.

Message lossMessage loss

Loss rate independent of experiment parameters.

(0 - 4%)

Messages are more likely to be highly delayed and reordered than lost.

14


Finding Design impactMessage reordering rateMessage reordering rate

Messages are reordered at a mean rate of 3.4%.

The protocol must be tolerant of message

reordering.

15


Finding Design impactMessage delayMessage delay

Delay is independent of intra-burst size.

Protocol does not need to regulate message

transmission.

Delay is highly correlated with the network interface and

time-of-day.

The protocol must tolerate highly variable delay.

16


0

0.2

0.4

0.6

0.8

1

1 10 100 1000

CD

F

Message delay (seconds)

Tethered cell phonePearl8820Bold

17

CDF of message delay


time-of-day.



0

0.2

0.4

0.6

0.8

1

1 10 100 1000

CD

F

Message delay (seconds)

Tethered cell phonePearl8820Bold

10

20

30

40

50

60

70

80

90

100

110

120

Midnight 3am 6am 9am Noon 3pm 6pm 9pm

Del

ay (s

econ

ds)

Time of the day

Mean delay 99% CIGlobal mean

17

CDF of message delay Mean message delay over a day


time-of-day.



Delay is independent of transmission index.

The protocol may be agnostic to the quantity of messages

transmitted.

0

20

40

60

80

100

120

140

0 10 20 30 40 50

Mea

n de

lay

(sec

onds

, 99%

CI)

Transmission index

18

Mean delay vs. transmission index


Despite high delays and reordering, messages are

received at a steady rate in batches.

The protocol may assume a relatively constant message

arrival rate.

19

0

0.2

0.4

0.6

0.8

1

1 10 100

CD

F

Inter-arrival time (seconds)

CDF of message inter-arrival time


Despite high delays and reordering, messages are

received at a steady rate in batches.

The protocol may assume a relatively constant message

arrival rate.

90% of messages delivered withinβ x mean

inter-arrival time

β≅3

19

0

0.2

0.4

0.6

0.8

1

1 10 100

CD

F

Inter-arrival time (seconds)

CDF of message inter-arrival time


• Preliminary experimentation

20



• Bi-directional SMS communication

20




• Increases transmission time and delay

20





• Message reordering (~45%)

20





• Message reordering (~45%)

• Losses are frequent (> 20%)

20


Outline

21



• Evaluation


Transport Protocol Design

• Goals

22



• Goals


22



• Goals


• Maximize throughput

22



• Goals


• Maximize throughput

• Flow control and error control

• Simplified version of NETBLT: SMS-TP

22


SMS-TP

23


SMS-TP

Fragment 1

23


SMS-TP

Selective ACK

Fragment 1

23


SMS-TP

Selective ACKWaits and avoids

bidirectional communication

Fragment 1

23


SMS-TP

Unfettered message

transmission

Selective ACKWaits and avoids

bidirectional communication

Fragment 1

23


SMS-TP

Single selective ACK tolerates

messagereordering

Unfettered message

transmission

Selective ACK

Final ACK

Waits and avoids bidirectional

communication

Fragment 1

23


SMS-TP

Selective ACK

When to send ACK?

Fragment 1

23


SMS-TP

Selective ACK

When to send ACK?

ACK

Fragment 1

23


SMS-TP

Selective ACK

When to send ACK?

ACK

Final ACK

Fragment 1

23


SMS-TP

Selective ACK

When to send ACK?

ACK

Final ACK

Fragment 1

23

Fragment 2


SMS-TP

Selective ACK

When to send ACK?

ACK

Final ACK

Final ACK

Fragment 1

23

Fragment 2


SMS-TP (receiver)

24


90% of messages are delivered within

β x mean inter-arrival time

SMS-TP (receiver)

24




SMS-TP (receiver)

EWA of inter-arrival

time

24




SMS-TP (receiver)


time

Message arrives within ACK timer

24




SMS-TP (receiver)


time


Final ACK

24




SMS-TP (receiver)


time


24

Has the databeen transferred?

Retransmit?


SMS-TP (sender)

25


SMS-TP (sender)

Setup delay(1 RTT)

25


SMS-TP (sender)

Setup delay(1 RTT)

ACK timer set to γ x setup

delay

25


SMS-TP (sender)

Setup delay(1 RTT)

Fragment 2


delay

ACK timer expires

25


SMS-TP (sender)

Setup delay(1 RTT)

Final ACK

Fragment 2


delay

Detects duplicate fragment,

retransmit final ACK

ACK timer expires

25


Outline

26



• Evaluation


Evaluation

• Implementation

27


Evaluation

• Implementation

• CLDC compliant Java library

27


Evaluation

• Implementation


• Free for download

27


Evaluation

• Implementation



• Evaluation

27


Evaluation

• Implementation



• Evaluation

• Field trial

27


Evaluation

• Implementation



• Evaluation

• Field trial

• Simulation

27


Field trial

28


Field trial

28

Sender

Receiver

Sender

Receiver


Field trial

28

Sender

Receiver

Sender

Receiver

SMS-TP SMS-TP

SMS-TP SMS-TP


Field trial

28

Stop-and-wait

Sender

Receiver

Sender

Receiver

SMS-TP

Stop-and-wait

SMS-TP

Stop-and-wait

SMS-TP

Stop-and-wait

SMS-TP


Field trial

28

Stop-and-wait

Optimal

Sender

Receiver

Sender

Receiver

SMS-TP

Stop-and-wait

Optimal

SMS-TP

Stop-and-wait

Optimal

SMS-TP

Stop-and-wait

Optimal

SMS-TP


Field trial

28

512 Bytes -32 KB

Stop-and-wait

512 Bytes - 32 KB

Optimal

Sender

Receiver

Sender

Receiver

SMS-TP

Stop-and-wait

Optimal

SMS-TP

Stop-and-wait

Optimal

SMS-TP

Stop-and-wait

Optimal

SMS-TP


29

Outperforms existing by a factor of two due to consolidation of ACKs

8.5% off calculated optimalPipelined transmission results in:

545% increase in performance


Sender

Receiver

29

Message overhead

0

50

100

150

200

250

300

350

400

450

500

512 B 1 KB 2 KB 4 KB 8 KB 16 KB 32 KB

Num

ber o

f SM

S m

essa

ges

Data size

Stop-and-waitSMS-TP (alpha = 0.5, beta = 3, gamma = 2)

Calculated optimal

Stop-and-wait

Optimal

SMS-TP

Stop-and-wait

Optimal

SMS-TPOutperforms existing by a factor of two due to consolidation of ACKs




Sender

Receiver

29

Data throughput

0

5

10

15

20

25

30

35

40

45

512 B 1 KB 2 KB 4 KB 8 KB 16 KB 32 KB

Thro

ughp

ut (b

ytes

/ se

cond

)

Data size


Calculated optimal

Message overhead

0

50

100

150

200

250

300

350

400

450

500

512 B 1 KB 2 KB 4 KB 8 KB 16 KB 32 KB

Num

ber o

f SM

S m

essa

ges

Data size


Calculated optimal

Stop-and-wait

Optimal

SMS-TP

Stop-and-wait

Optimal

SMS-TPOutperforms existing by a factor of two due to consolidation of ACKs




Simulation

30


Simulation

• Study may not reflect conditions of a developing region (Zerfos’06)

30


Simulation


• Loss rate

30


Simulation


• Loss rate

• Delay

30


Simulation


• Loss rate

• Delay

Sender

Receiver

30


Simulation


• Loss rate

• Delay

SMS-TP SMS-TPSender

Receiver

30


Simulation


• Loss rate

• Delay

SMS-TP SMS-TPSender

ReceiverSMS-SIM

30


Simulation


• Loss rate

• Delay

SMS-TP SMS-TPSender

ReceiverSMS-SIM

Loss rate: 1% - 10%

30


Simulation


• Loss rate

• Delay

SMS-TP SMS-TPSender

ReceiverSMS-SIM

Delay: 30 seconds - 3 minutes

30


SMS-TP SMS-TPSender

Receiver

31

SMS-SIM

A loss rate increase has statistically insignificant impact on throughput

6 fold increase in delay, 50% decrease in throughput

50% better than stop-and-wait in a well provisioned network

Delay and loss have a negligible impact on message overhead- except under high loss and high delay situations


0

50

100

150

200

250

20 40 60 80 100 120 140 160 180

Num

ber o

f SM

S m

essa

ges

Delay (seconds)

Loss rate = 0.01Loss rate = 0.05

Loss rate = 0.1

Message overhead

SMS-TP SMS-TPSender

Receiver

31

SMS-SIM






0

50

100

150

200

250

20 40 60 80 100 120 140 160 180

Num

ber o

f SM

S m

essa

ges

Delay (seconds)


Loss rate = 0.1

Message overhead

0

5

10

15

20

25

30

35

40

45

20 40 60 80 100 120 140 160 180

Thro

ughp

ut (b

ytes

/ se

cond

)

Delay (seconds)


Loss rate = 0.1

Data throughput

SMS-TP SMS-TPSender

Receiver

31

SMS-SIM






Summary

• Studied the channel characteristics of SMS from the perspective of mobile devices sending bursts of messages

32


Summary



32


Summary



• Reduces message overhead by 50%

32


Summary



• Reduces message overhead by 50%

• Increases throughput by as much as 545%

32


Questions?

33

Why not TCP?

- Significant delays- Messages rarely lost- Reordering is common- Does not suffer from congestion drops


characterizing the transport behaviour of the short

Documents