Can We Multiplex ACKs without Harming the Performance of TCP?

25
Can We Multiplex ACKs without Harming the Performance of TCP? CCNC 2014, The 11 Th Annual IEEE Consumer Communications & Networking Conference January 10-13 Las Vegas, Nevada USA Jose Saldana, Julián Fernández-Navajas, José Ruiz-Mas

description

Jose Saldana, Julian Fernandez-Navajas, Jose Ruiz-Mas "Can We Multiplex ACKs without Harming the Performance of TCP?," Consumer Communications and Networking Conference, CCNC 2014. Las Vegas, January 10, 2014, pp 921-922. ISBN 978-1-4799-2356-4.

Transcript of Can We Multiplex ACKs without Harming the Performance of TCP?

Page 1: Can We Multiplex ACKs without Harming the Performance of TCP?

Can We Multiplex ACKs

without Harming the

Performance of TCP?

CCNC 2014, The 11Th Annual IEEE Consumer Communications & Networking Conference

January 10-13 Las Vegas, Nevada USA

Jose Saldana, Julián Fernández-Navajas, José Ruiz-Mas

Page 2: Can We Multiplex ACKs without Harming the Performance of TCP?

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Index

1. Introduction

2. Tests and results

3. Conclusions

Page 3: Can We Multiplex ACKs without Harming the Performance of TCP?

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Index

1. Introduction

2. Tests and results

3. Conclusions

Page 4: Can We Multiplex ACKs without Harming the Performance of TCP?

Introduction

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Increase of emerging real-time services

They use small packets

This is modifying the traffic mix present

on the Internet

Inefficiency of the packets

IPv6 makes the problem even worse

VoIP packets

Page 5: Can We Multiplex ACKs without Harming the Performance of TCP?

Introduction

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

TCRTP (RFC4170) improves the efficiency

of VoIP. It uses three layers:

Header compression

Multiplexing

Tunneling

IP network

MUX DEMUX .

.

.

.

.

.

RTP RTP multiplexing RTP

Page 6: Can We Multiplex ACKs without Harming the Performance of TCP?

Introduction

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Advantage: Bandwidth and pps savings

At the cost of an additional multiplexing delay

Inter - pkt time

. . .

. . .

Native VoIP

traffic

Optimized

traffic

Inter - pkt time Inter - pkt time

. . .

. . .

. . .

. . .

. . .

. . .

Page 7: Can We Multiplex ACKs without Harming the Performance of TCP?

Introduction

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

TCM-TF*: Proposal for multiplexing other traffic

flows, including UDP (non-RTP) and TCP

*draft-saldana-tsvwg-tcmtf-05

IP IP IP

No compr. / ROHC / IPHC / ECRTP

PPPMux / Other

GRE / L2TP

IP

Compression layer

Multiplexing layer

Tunneling layer

a) TCRTP

Network Protocol

UDP

RTP

payload

UDPTCP

payloadpayload

ECRTP

PPPMux

L2TP

IP

IP

UDP

RTP

payload

b) TCMTF

MPLS

Page 8: Can We Multiplex ACKs without Harming the Performance of TCP?

Introduction

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

TCP video traffic: 69% of all consumer Internet traffic

in 2017.

When downloading a video, a computer may

generate some hundreds of ACKs per second, during

some tens of seconds.

In some scenarios (e.g. the aggregation network of an

operator) high numbers of long-term flows of ACKs

share a common path.

Header compression ratio of ACKs: from 40 to 7 or 8

bytes (savings of 80%).

Page 9: Can We Multiplex ACKs without Harming the Performance of TCP?

Introduction

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Is it a good idea to compress and multiplex these

flows?

Would the multiplexing delay degrade the

performance of TCP?

Sawtooth-shaped delay

PE

PE time

Added

delay

Page 10: Can We Multiplex ACKs without Harming the Performance of TCP?

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Index

1. Introduction

2. Tests and results

3. Conclusions

Page 11: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Dumbbell scenario in ns2

A sawtooth-shaped delay is added to the ACKs B-B’

What is the effect? We use TCP Tahoe (the most

basic one) in order to more clearly see the effect

First tests: separate A-A’ and B-B’

A

B

N M

A’

B’

O

P ACK mux (PE)

FTP

FTP

ACK

Page 12: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

0

2

4

6

8

10

30 35 40 45 50 55 60Ban

dw

idth

[M

bp

s]

simulated time [s]

Throughput (RTT = 80 ms)

0

2

4

6

8

10

30 35 40 45 50 55 60Ba

nd

wid

th [

Mb

ps]

simulated time [s]

Throughput (RTT = 80 ms, mux period 50 ms)

Page 13: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

0

2

4

6

8

10

30 35 40 45 50 55 60Ban

dw

idth

[M

bp

s]

simulated time [s]

Throughput (RTT = 80 ms)

0

2

4

6

8

10

30 35 40 45 50 55 60Ba

nd

wid

th [

Mb

ps]

simulated time [s]

Throughput (RTT = 80 ms, mux period 50 ms)

avg 9.24 Mbps window reset

every ~7 sec

avg 8.04 Mbps

(12% reduction)

window reset

every ~9,5 sec

ACKs arrive

in bursts

Page 14: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

0

50

100

150

200

250

0 10 20 30 40 50 60 70 80

Win

do

w s

ize

simulation time [s]

Window size

no PE

PE=50ms

Page 15: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

0

20

40

60

80

100

120

140

Win

do

w s

ize

Window evolution. One period

no PE

PE=50ms

Slow start

ends later

Window size

increases

more slowly

The period

between

window

resets

increases

Page 16: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Second tests: A-A’ and B-B’ share the bottleneck

Are multiplexed flows in clear disadvantage?

We will use four different TCP variants:

Tahoe

Reno

New Reno

SACK

Results: Throughput difference between

multiplexed and non-multiplexed flows

A

B

N M

A’

B’

O

P ACK mux (PE)

FTP

FTP

ACK

Page 17: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Results: Throughput difference between

multiplexed and non-multiplexed flows

Multiplexing Period PE [ms]

TCP 5 10 15 20 25

Tahoe 4.91 % 10.05 % 31.67 % 7.88 % 49.74 %

Reno 5.95 % 17.78 % 48.62 % 24.29 % 61.92 %

New

Reno 4.82 % 12.95 % 30.52 % 16.70 % 52.87 %

SACK 2.27 % 12.70 % 20.62 % 14.75 % 50.90 %

Page 18: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Results: Throughput difference between

multiplexed and non-multiplexed flows

Multiplexing Period PE [ms]

TCP 5 10 15 20 25

Tahoe 4.91 % 10.05 % 31.67 % 7.88 % 49.74 %

Reno 5.95 % 17.78 % 48.62 % 24.29 % 61.92 %

New

Reno 4.82 % 12.95 % 30.52 % 16.70 % 52.87 %

SACK 2.27 % 12.70 % 20.62 % 14.75 % 50.90 %

With PE=5 ms

the difference

is small

Page 19: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Results: Throughput difference between

multiplexed and non-multiplexed flows

Multiplexing Period PE [ms]

TCP 5 10 15 20 25

Tahoe 4.91 % 10.05 % 31.67 % 7.88 % 49.74 %

Reno 5.95 % 17.78 % 48.62 % 24.29 % 61.92 %

New

Reno 4.82 % 12.95 % 30.52 % 16.70 % 52.87 %

SACK 2.27 % 12.70 % 20.62 % 14.75 % 50.90 %

With PE=10 ms

the difference

becomes higher

Page 20: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Results: Throughput difference between

multiplexed and non-multiplexed flows

Multiplexing Period PE [ms]

TCP 5 10 15 20 25

Tahoe 4.91 % 10.05 % 31.67 % 7.88 % 49.74 %

Reno 5.95 % 17.78 % 48.62 % 24.29 % 61.92 %

New

Reno 4.82 % 12.95 % 30.52 % 16.70 % 52.87 %

SACK 2.27 % 12.70 % 20.62 % 14.75 % 50.90 %

Below 10 ms the

difference may

become

unacceptable

Page 21: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

0

2

4

6

8

10

900 910 920 930 940 950 960 970 980 990 1000

Th

rou

gh

pu

t [M

bp

s]

Simulation time

Throughput (SACK) no PE

PE=5 ms

Page 22: Can We Multiplex ACKs without Harming the Performance of TCP?

Tests and Results

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

0

2

4

6

8

10

900 910 920 930 940 950 960 970 980 990 1000

Th

rou

gh

pu

t [M

bp

s]

Simulation time

Throughput (Reno) no PE

PE=25 ms

Page 23: Can We Multiplex ACKs without Harming the Performance of TCP?

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Index

1. Introduction

2. Tests and results

3. Conclusions

Page 24: Can We Multiplex ACKs without Harming the Performance of TCP?

Conclusions

Can We Multiplex ACKs without Harming the Performance of TCP? - CCNC 2014

Suitability of traffic optimization, based on header

compression and multiplexing, to the flows of ACKs

The expected bandwidth savings are huge because of

the absence of payload

Counterpart: throughput reduction when an

optimized flow shares a bottleneck with a non-

optimized one

The impairments can be maintained in tolerable

limits, by setting an upper bound on the period

Further study this trade-off between bandwidth

saving and throughput reduction

Page 25: Can We Multiplex ACKs without Harming the Performance of TCP?

Thank you very much!

Jose Saldana, Julián Fernández-Navajas, José Ruiz-Mas