Implementing CAIA Delay-Gradient in Linux Kenneth Klette Jonassen - - PowerPoint PPT Presentation

▶
implementing caia delay gradient in linux
SMART_READER_LITE
LIVE PREVIEW

Implementing CAIA Delay-Gradient in Linux Kenneth Klette Jonassen - - PowerPoint PPT Presentation

Apr 20, 2023 156 likes •934 views

Introduction Tests Recent advances Future work and contributions Extra slides Implementing CAIA Delay-Gradient in Linux Kenneth Klette Jonassen Department of Informatics University of Oslo Introduction Tests Recent advances Future work

slide-1
SLIDE 1

Introduction Tests Recent advances Future work and contributions Extra slides

Implementing CAIA Delay-Gradient in Linux

Kenneth Klette Jonassen Department of Informatics University of Oslo

slide-2
SLIDE 2

Introduction Tests Recent advances Future work and contributions Extra slides

Outline

Introduction Thesis subject Thesis outline Tests Recent advances Future work and contributions Future work Linux contributions Early access Extra slides

slide-3
SLIDE 3

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

slide-4
SLIDE 4

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

slide-5
SLIDE 5

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

slide-6
SLIDE 6

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

◮ CDG is influenced by previous congestion controls

slide-7
SLIDE 7

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

◮ CDG is influenced by previous congestion controls

◮ Our contribution is to make CDG work for Linux

slide-8
SLIDE 8

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

◮ CDG is influenced by previous congestion controls

◮ Our contribution is to make CDG work for Linux ◮ The optimal outcome is to get our CDG implementation

accepted into mainline Linux

slide-9
SLIDE 9

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

◮ CDG is influenced by previous congestion controls

◮ Our contribution is to make CDG work for Linux ◮ The optimal outcome is to get our CDG implementation

accepted into mainline Linux

◮ Keep it maintained for future versions of Linux

slide-10
SLIDE 10

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

◮ CDG is influenced by previous congestion controls

◮ Our contribution is to make CDG work for Linux ◮ The optimal outcome is to get our CDG implementation

accepted into mainline Linux

◮ Keep it maintained for future versions of Linux ◮ Make CDG available to a broad audience of users

slide-11
SLIDE 11

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis subject

◮ The goal of this thesis is to implement CAIA

Delay-Gradient (CDG) in Linux

◮ CDG is the result of previous work done at Swinburne

University’s Centre for Advanced Internet Architectures

◮ CDG is influenced by previous congestion controls

◮ Our contribution is to make CDG work for Linux ◮ The optimal outcome is to get our CDG implementation

accepted into mainline Linux

◮ Keep it maintained for future versions of Linux ◮ Make CDG available to a broad audience of users ◮ See benefits to real-world applications outside academia

slide-12
SLIDE 12

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

slide-13
SLIDE 13

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control

slide-14
SLIDE 14

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control

slide-15
SLIDE 15

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal

slide-16
SLIDE 16

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

slide-17
SLIDE 17

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

slide-18
SLIDE 18

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux

slide-19
SLIDE 19

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux ◮ Shortcomings of Linux’ RTT measurements

slide-20
SLIDE 20

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux ◮ Shortcomings of Linux’ RTT measurements ◮ ... and how these were addressed.

slide-21
SLIDE 21

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux ◮ Shortcomings of Linux’ RTT measurements ◮ ... and how these were addressed. ◮ Valid measurements are vital to CDG’s performance

slide-22
SLIDE 22

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux ◮ Shortcomings of Linux’ RTT measurements ◮ ... and how these were addressed. ◮ Valid measurements are vital to CDG’s performance

◮ Testbed setup

slide-23
SLIDE 23

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux ◮ Shortcomings of Linux’ RTT measurements ◮ ... and how these were addressed. ◮ Valid measurements are vital to CDG’s performance

◮ Testbed setup

◮ Mitigating hidden variables that could distort or bias results

slide-24
SLIDE 24

Introduction Tests Recent advances Future work and contributions Extra slides

Thesis outline

◮ Internet congestion control

◮ Transmission Control Protocol (TCP) ◮ CDG is a TCP congestion control ◮ Network delays ◮ CDG is a delay-based congestion control ◮ Why delay works as a congestion signal ◮ When delay does not work as a congestion signal

◮ Implementing CDG in Linux

◮ Preliminaries for congestion control development in Linux ◮ Shortcomings of Linux’ RTT measurements ◮ ... and how these were addressed. ◮ Valid measurements are vital to CDG’s performance

◮ Testbed setup

◮ Mitigating hidden variables that could distort or bias results

◮ Evaluation of homogeneous capacity sharing

slide-25
SLIDE 25

Introduction Tests Recent advances Future work and contributions Extra slides

Experiment testbed

Linux router

Switch

Linux receiver FreeBSD sender Linux sender Gigabit ethernet Management

◮ Dedicated gigabit links between the router and hosts ◮ Management network runs over seperate network interfaces ◮ Using HTB to shape traffic to 10 Mbps ◮ Using netem to emulate RTT of 40 ms or 70 ms ◮ 84p FIFO queue with Tail drop

slide-26
SLIDE 26

Introduction Tests Recent advances Future work and contributions Extra slides

vs. CDG CUBIC 40 ms RTT 10 Mbps bottleneck 84p FIFO queue with Tail drop

slide-27
SLIDE 27

Introduction Tests Recent advances Future work and contributions Extra slides

flent

http://www.flent.org/

(formerly known as netperf-wrapper)

slide-28
SLIDE 28

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 2 4 6 8 10 12 Mbits/s

cdg

Upload Ping (ms)

20 40 60 80 100 120 140 ms 50 100 150 200 250 300 Time (s) 2 4 6 8 10 12 Mbits/s

cubic

Upload Ping (ms)

20 40 60 80 100 120 140 ms

Single TCP upload stream w/ping Bandwidth and ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:13:44.876320 - Length/step: 300s/0.20s

slide-29
SLIDE 29

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 2 4 6 8 10 Mbits/s

cdg

Stream 1 Stream 2

50 100 150 200 250 300 Time (s) 2 4 6 8 10 Mbits/s

cubic

Stream 1 Stream 2

Two TCP upload streams; 2nd stream started delayed Bandwidth plot, 5 second smoothed sample size

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:24:46.627611 - Length/step: 300s/0.20s

slide-30
SLIDE 30

Introduction Tests Recent advances Future work and contributions Extra slides

627611.cdg 523368.cubic 1 2 3 4 Mean bandwidth (MBits/s)

Stream 1 Stream 2

Two TCP upload streams; 2nd stream started delayed Summarising combined tests

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:24:46.627611 - Length/step: 300s/0.20s

slide-31
SLIDE 31

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 2 4 6 8 10 Mbits/s

cdg

Download Ping (ms)

20 40 60 80 100 120 Latency (ms) 50 100 150 200 250 300 Time (s) 2 4 6 8 10 Mbits/s

cubic

Download Ping (ms)

20 40 60 80 100 120 Latency (ms)

8 down - dslreports dsl test equivalent Total bandwidth and average ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:46:42.323957 - Length/step: 300s/0.20s

slide-32
SLIDE 32

Introduction Tests Recent advances Future work and contributions Extra slides

Download Ping (ms) 2 4 6 8 10 12 Mbits/s

cdg cubic

20 40 60 80 100 120 140 Latency (ms)

8 down - dslreports dsl test equivalent Box plot of totals

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:46:42.323957 - Length/step: 300s/0.20s

slide-33
SLIDE 33

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 2 4 6 8 10 Mbits/s

cdg

Download Upload Ping (ms)

20 40 60 80 100 120 Latency (ms) 50 100 150 200 250 300 Time (s) 2 4 6 8 10 Mbits/s

cubic

Download Upload Ping (ms)

20 40 60 80 100 120 Latency (ms)

Realtime Response Under Load Total bandwidth and average ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:57:46.372348 - Length/step: 300s/0.20s

slide-34
SLIDE 34

Introduction Tests Recent advances Future work and contributions Extra slides

Download Upload Ping (ms) 2 4 6 8 10 12 14 16 Mbits/s

cdg cubic

20 40 60 80 100 120 140 Latency (ms)

Realtime Response Under Load Box plot of totals

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:57:46.372348 - Length/step: 300s/0.20s

slide-35
SLIDE 35

Introduction Tests Recent advances Future work and contributions Extra slides

20 40 60 80 100 120 140 ms 5 10 15 20 Mbits/s cdg cubic

Realtime Response Under Load Ellipsis plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 00:57:46.372348 - Length/step: 300s/0.20s

slide-36
SLIDE 36

Introduction Tests Recent advances Future work and contributions Extra slides

vs. CDG CUBIC 40 ms RTT 100 Mbps bottleneck 840p FIFO queue with Tail drop

slide-37
SLIDE 37

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 20 40 60 80 100 Mbits/s

cdg

Upload Ping (ms)

10 20 30 40 50 60 70 80 ms 50 100 150 200 250 300 Time (s) 20 40 60 80 100 Mbits/s

cubic

Upload Ping (ms)

10 20 30 40 50 60 70 80 ms

Single TCP upload stream w/ping Bandwidth and ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 01:56:42.773456 - Length/step: 300s/0.20s

slide-38
SLIDE 38

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 20 40 60 80 Mbits/s

cdg

Stream 1 Stream 2

50 100 150 200 250 300 Time (s) 20 40 60 80 Mbits/s

cubic

Stream 1 Stream 2

Two TCP upload streams; 2nd stream started delayed Bandwidth plot, 5 second smoothed sample size

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:07:46.457561 - Length/step: 300s/0.20s

slide-39
SLIDE 39

Introduction Tests Recent advances Future work and contributions Extra slides

457561.cdg 384635.cubic 10 20 30 40 Mean bandwidth (MBits/s)

Stream 1 Stream 2

Two TCP upload streams; 2nd stream started delayed Summarising combined tests

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:07:46.457561 - Length/step: 300s/0.20s

slide-40
SLIDE 40

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 20 40 60 80 100 Mbits/s

cdg

Download Ping (ms)

20 40 60 80 100 120 140 Latency (ms) 50 100 150 200 250 300 Time (s) 20 40 60 80 100 Mbits/s

cubic

Download Ping (ms)

20 40 60 80 100 120 140 Latency (ms)

8 down - dslreports dsl test equivalent Total bandwidth and average ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:36:10.066039 - Length/step: 300s/0.20s

slide-41
SLIDE 41

Introduction Tests Recent advances Future work and contributions Extra slides

Download Ping (ms) 20 40 60 80 100 120 Mbits/s

cdg cubic

20 40 60 80 100 120 140 Latency (ms)

8 down - dslreports dsl test equivalent Box plot of totals

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:36:10.066039 - Length/step: 300s/0.20s

slide-42
SLIDE 42

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 20 40 60 80 100 Mbits/s

cdg

Download Upload Ping (ms)

20 40 60 80 100 120 Latency (ms) 50 100 150 200 250 300 Time (s) 20 40 60 80 100 Mbits/s

cubic

Download Upload Ping (ms)

20 40 60 80 100 120 Latency (ms)

Realtime Response Under Load Total bandwidth and average ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:47:14.909773 - Length/step: 300s/0.20s

slide-43
SLIDE 43

Introduction Tests Recent advances Future work and contributions Extra slides

Download Upload Ping (ms) 20 40 60 80 100 120 Mbits/s

cdg cubic

20 40 60 80 100 120 Latency (ms)

Realtime Response Under Load Box plot of totals

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:47:14.909773 - Length/step: 300s/0.20s

slide-44
SLIDE 44

Introduction Tests Recent advances Future work and contributions Extra slides

20 40 60 80 100 120 140 ms 50 100 150 200 Mbits/s cdg cubic

Realtime Response Under Load Ellipsis plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:47:14.909773 - Length/step: 300s/0.20s

slide-45
SLIDE 45

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 20 40 60 80 100 120 Mbits/s

cdg

Download Upload Ping (ms)

20 40 60 80 100 120 Latency (ms) 50 100 150 200 250 300 Time (s) 20 40 60 80 100 120 Mbits/s

cubic

Download Upload Ping (ms)

20 40 60 80 100 120 Latency (ms)

50 up vs 1 down - exclusively Best Effort Total bandwidth and average ping plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 03:32:07.805839 - Length/step: 300s/0.20s

slide-46
SLIDE 46

Introduction Tests Recent advances Future work and contributions Extra slides

50 100 150 200 250 300 10 20 30 40 50 60 70 Mbits/s TCP download

CDG BE Cubic BE CDG CS1 Cubic CS1 Avg

50 100 150 200 250 300 10 20 30 40 50 60 70 Mbits/s TCP upload

CDG BE Cubic BE CDG CS1 Cubic CS1 Avg

50 100 150 200 250 300 Time (s) 20 40 60 80 100 120 Latency (ms) Ping (ms)

UDP CS6 UDP CS1 UDP BE ICMP Avg

Cubic VS CDG Download, upload, ping (unscaled versions)

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:30:38.827305 - Length/step: 300s/0.20s

slide-47
SLIDE 47

Introduction Tests Recent advances Future work and contributions Extra slides

CDG BE Cubic BE CDG CS1 Cubic CS1 Avg 10 20 30 40 50 60 70 Mbits/s

Cubic VS CDG Download bandwidth box plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:30:38.827305 - Length/step: 300s/0.20s

slide-48
SLIDE 48

Introduction Tests Recent advances Future work and contributions Extra slides

CDG BE Cubic BE CDG CS1 Cubic CS1 Avg 10 20 30 40 50 60 70 Mbits/s

Cubic VS CDG Upload bandwidth box plot

Local/remote: bestemor/10.0.1.2 - Time: 2015-06-15 02:30:38.827305 - Length/step: 300s/0.20s

slide-49
SLIDE 49

Introduction Tests Recent advances Future work and contributions Extra slides

CDG has parameters to tune its backoff behavior (p.61)

G = 1 G = 2 G = 3 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 2 × 0 ms delay 2 × 10 ms delay 2 × 20 ms delay 2 × 35 ms delay 40 80 40 80 40 80

Queueing delay (msecs) Empirical CDF

Implementation FreeBSD Linux

Homogeneous capacity sharing – CDG G ∈ {1, 2, 3} Delay ∈ {0, 10, 20, 35} Figure 4.7: Queueing delay distributions, as measured by the sender and the receiver in cooperation.

slide-50
SLIDE 50

Introduction Tests Recent advances Future work and contributions Extra slides

Recent advances

slide-51
SLIDE 51

Introduction Tests Recent advances Future work and contributions Extra slides

CDG is now a part of net-next

slide-52
SLIDE 52

Introduction Tests Recent advances Future work and contributions Extra slides

CDG was featured on lwn.net

. . .

slide-53
SLIDE 53

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

slide-54
SLIDE 54

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay

slide-55
SLIDE 55

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay ◮ CDG can exit slow start using its probabilistic backoff

slide-56
SLIDE 56

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay ◮ CDG can exit slow start using its probabilistic backoff ◮ HyStart can exit slow start deterministically

slide-57
SLIDE 57

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay ◮ CDG can exit slow start using its probabilistic backoff ◮ HyStart can exit slow start deterministically

◮ CUBIC’s congestion avoidance

◮ Faster rate increase to help CDG level with CUBIC

slide-58
SLIDE 58

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay ◮ CDG can exit slow start using its probabilistic backoff ◮ HyStart can exit slow start deterministically

◮ CUBIC’s congestion avoidance

◮ Faster rate increase to help CDG level with CUBIC

◮ Parameters for using CDG in different environments

◮ Defaults are magic numbers

slide-59
SLIDE 59

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay ◮ CDG can exit slow start using its probabilistic backoff ◮ HyStart can exit slow start deterministically

◮ CUBIC’s congestion avoidance

◮ Faster rate increase to help CDG level with CUBIC

◮ Parameters for using CDG in different environments

◮ Defaults are magic numbers

◮ One-way delay measurements

◮ Avoid conflating delay in forward and reverse paths, e.g.,

simultaneous download and upload

slide-60
SLIDE 60

Introduction Tests Recent advances Future work and contributions Extra slides

Future work

◮ Hybrid Slow start (HyStart)

◮ Avoiding slow start overshoot avoids loss and delay ◮ CDG can exit slow start using its probabilistic backoff ◮ HyStart can exit slow start deterministically

◮ CUBIC’s congestion avoidance

◮ Faster rate increase to help CDG level with CUBIC

◮ Parameters for using CDG in different environments

◮ Defaults are magic numbers

◮ One-way delay measurements

◮ Avoid conflating delay in forward and reverse paths, e.g.,

simultaneous download and upload

◮ Hardware RTT measurements

◮ Possible with some recent adapters

slide-61
SLIDE 61

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot

slide-62
SLIDE 62

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

slide-63
SLIDE 63

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

◮ “Linux TCP/Reno behavior with large DropTail Queue”

(rite-research e-mail)

◮ Should we add a global PRR on/off knob in Linux?

slide-64
SLIDE 64

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

◮ “Linux TCP/Reno behavior with large DropTail Queue”

(rite-research e-mail)

◮ Should we add a global PRR on/off knob in Linux?

◮ ... and tcp_cdg now has HyStart!

slide-65
SLIDE 65

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

◮ “Linux TCP/Reno behavior with large DropTail Queue”

(rite-research e-mail)

◮ Should we add a global PRR on/off knob in Linux?

◮ ... and tcp_cdg now has HyStart!

◮ Previous HyStart code is tightly tangled into tcp_cubic

slide-66
SLIDE 66

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

◮ “Linux TCP/Reno behavior with large DropTail Queue”

(rite-research e-mail)

◮ Should we add a global PRR on/off knob in Linux?

◮ ... and tcp_cdg now has HyStart!

◮ Previous HyStart code is tightly tangled into tcp_cubic ◮ We have an opportunity to improve HyStart ◮ tcp_cubic is the default, and is sensitive to change ◮ tcp_cdg is a new module

slide-67
SLIDE 67

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

◮ “Linux TCP/Reno behavior with large DropTail Queue”

(rite-research e-mail)

◮ Should we add a global PRR on/off knob in Linux?

◮ ... and tcp_cdg now has HyStart!

◮ Previous HyStart code is tightly tangled into tcp_cubic ◮ We have an opportunity to improve HyStart ◮ tcp_cubic is the default, and is sensitive to change ◮ tcp_cdg is a new module ◮ CDG’s HyStart uses Hayes’ Delayed ACK filter ◮ Helps avoid spurious slow start exits

slide-68
SLIDE 68

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start

◮ tcp_cubic has HyStart

◮ Important to avoid slow start overshoot ◮ Alleviates an issue that could be caused by Proportional

Rate Reduction (PRR)

◮ “Linux TCP/Reno behavior with large DropTail Queue”

(rite-research e-mail)

◮ Should we add a global PRR on/off knob in Linux?

◮ ... and tcp_cdg now has HyStart!

◮ Previous HyStart code is tightly tangled into tcp_cubic ◮ We have an opportunity to improve HyStart ◮ tcp_cubic is the default, and is sensitive to change ◮ tcp_cdg is a new module ◮ CDG’s HyStart uses Hayes’ Delayed ACK filter ◮ Helps avoid spurious slow start exits ◮ CDG’s HyStart uses a microsecond clock ◮ Increases accuracy for high speeds

slide-69
SLIDE 69

Introduction Tests Recent advances Future work and contributions Extra slides

Contributions (patches) to Linux

Commit Description Linux 3725a26 pkt_sched: fq: avoid hang when ... 3.19 – pkt_sched: fq: avoid artificial bursts ... – 932eb76 tcp: use RTT from SACK for CC 4.0 3d0d26c tcp: fix bogus RTT when retrans... 4.1 196da97 tcp: move struct tcp_sacktag_state to ... (4.2) 31231a8 tcp: improve RTT from SACK for CC (4.2) 138998f tcp: invoke pkts_acked hook on every ACK (4.2) 7782ad8 tcp: export tcp_enter_cwr() (4.2) 2b0a8c9 tcp: add CDG congestion control (4.2) 758f0d4 tcp: cdg: use div_u64() (4.2) Patches for Linux 4.2 are in net-next, where they remain until Linux 4.1 is released and the merge window for Linux 4.2 opens.

slide-70
SLIDE 70

Introduction Tests Recent advances Future work and contributions Extra slides

Early access

Clone Linux’ net-next git tree:

git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-next.git

Select to compile tcp_cdg, e.g., using make menuconfig: . . .

slide-71
SLIDE 71

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

slide-72
SLIDE 72

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

◮ Setups using pacing must disable ACK train detection

slide-73
SLIDE 73

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

◮ Setups using pacing must disable ACK train detection

◮ The ACK train algorithm probes the Bandwidth Delay

Product by guessing the sum of inter-packet arrival times

slide-74
SLIDE 74

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

◮ Setups using pacing must disable ACK train detection

◮ The ACK train algorithm probes the Bandwidth Delay

Product by guessing the sum of inter-packet arrival times

◮ It assumes that packets are sent in bursts

slide-75
SLIDE 75

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

◮ Setups using pacing must disable ACK train detection

◮ The ACK train algorithm probes the Bandwidth Delay

Product by guessing the sum of inter-packet arrival times

◮ It assumes that packets are sent in bursts ◮ Pacing spaces the inter-packet transmission times

slide-76
SLIDE 76

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

◮ Setups using pacing must disable ACK train detection

◮ The ACK train algorithm probes the Bandwidth Delay

Product by guessing the sum of inter-packet arrival times

◮ It assumes that packets are sent in bursts ◮ Pacing spaces the inter-packet transmission times

◮ Can still use the delay threshold algorithm

slide-77
SLIDE 77

Introduction Tests Recent advances Future work and contributions Extra slides

Hybrid Slow start with pacing

◮ Hybrid Slow start consists of two algorithms to help probe

the slow start threshold:

◮ ACK train detection ◮ Delay threshold detection

◮ Setups using pacing must disable ACK train detection

◮ The ACK train algorithm probes the Bandwidth Delay

Product by guessing the sum of inter-packet arrival times

◮ It assumes that packets are sent in bursts ◮ Pacing spaces the inter-packet transmission times

◮ Can still use the delay threshold algorithm

◮ Set parameter hystart_detect=2 for tcp_cubic and

tcp_cdg to selectively enable the delay threshold algorithm.