Drongo Speeding Up CDNs with Subnet Assimilation from the Client

Authors:Marc Anthony WarriorUri KlarmanMarcel FloresAleksandar Kuzmanovic

CoNEXT ‘17Incheon, South KoreaCDN & Caching Session

Bird’s Eye View● What is Drongo?

● Why we need Drongo

● Performance Analysis

● Thoughts & Conclusions

● Questions

22

What is Drongo?

3

What is Drongo?

It’s a bird!

4

What is Drongo?

It’s a bird!

5

What is Drongo?

It’s a bird!

6

What is Drongo?

It’s a system that allows end-users to enhance the QoS (quality of service) they get from Content Distribution Networks (CDNs)

7

What is Drongo?

It’s a system that allows end-users to enhance the QoS (quality of service) they get from Content Distribution Networks (CDNs)

(in this talk, QoS = latency)

8

Why Latency?

9

Why Latency?● Latency is time

10

Why Latency?● Latency is time

● Latency is money ○ Google (Marissa Mayer), Amazon (Greg Linden)

■ Web 2.0 Summet, glinden.blogspot.com

11

Why Latency?● Latency is time

● Latency is money ○ Google (Marissa Mayer), Amazon (Greg Linden)

■ Web 2.0 Summet, glinden.blogspot.com

● Latency is the bottom line ○ “What we have found running our applications at

Google is that latency is as important, or more important, for our applications than relative bandwidth,” Amin Vahdat (Google)

12

Drongo helps you (the end user)

lower your own latency!

13

14

Drongo’s Effect on Latency

GoogleAmazon

AlibabaCDNetworks

ChinaNetCtr

CubeCDN

15

Drongo’s Effect on Latency

GoogleAmazon

AlibabaCDNetworks

ChinaNetCtr

CubeCDN

ONLY client-side changes

Example Scenario

16

Provider wants to serve client17

Client is far18

CDN = more replica locations19

DNS Redirection

Which replica serves the client?

20

Choose the “closest” server21

Choose the “closest” server

This choice is nontrivial!

22

Often Suboptimal Choices!23

24

Maybe just a far LDNS...

[Chen - SigComm ’15; Huang - SigComm CCR ‘12; Alzoubi - WWW ‘13; Rula - SigComm ‘14 …]

Ordinary DNS Query

25

DNS Query LDNS IP

Somewhere in California

EDNS0 Client-Subnet extension (ECS)

26

DNS Query LDNS IP

Somewhere in California Actually somewhere in New York

Client Subnet

27

(ECS User)

We used ECS:

This still happens

28

(ECS User)

We used ECS:

This still happens

29

… frequently

(ECS User)

We used ECS:

Really? ...

30

Really? ...

YES! We measured it!

31

How did we measure it?

32

How did we measure it?

33

Find subnets directed to different replicas

Subnet Assimilation

34

DNS Query LDNS IPClient Subnet

Subnet Assimilation

35

DNS Query LDNS IPClient SubnetOther Subnet

How did we measure it?

36

Find subnets directed to different replicas

Perform pings and downloads to each replica

How did we measure it?

37

Find subnets directed to different replicas

Perform pings and downloads to each replica

Identify which subnet resulted in the “best” replica

1. Get “Default” Choice38

(use client’s own subnet for ECS)

2. Traceroute to default choice39

3. Get Hop Subnet Choices40

(use hops’ subnets for ECS)

4. Measure Latencies41

4. Measure Latencies42

Steps 1-4: a “trial”

Latency Ratio 43

10.6

1.4

Normalize to default choice’s RTT

We’re looking for this 44

10.6

1.4

Valley = better choice from hop subnet

100 ms

0 ms

RTT: client to replica

traceroute

replica choice for subnet

45

Valley = better choice from hop subnet

100 ms

0 ms

RTT: client to replica

traceroute

replica choice for subnet

46

PlanetLab Sees Valleys!

47

PlanetLab Sees Valleys!

48

PlanetLab Sees Valleys!

● Google: 20.24%● Amazon: 14.02%● Alibaba: 33.68%● CDNetworks: 15.61%● ChinaNetCenter: 27.42%● CubeCDN: 38.58% Room for improvement!

49

5.50

5. Use best subnet for ECS51

5. Use best subnet for ECS52

Get best mapping!

Are Valleys Predictable?

● Trials are not “fast”

53

Are Valleys Predictable?

● Trials are not “fast”● We want valleys “on the fly”

54

Are Valleys Predictable?

● Trials are not “fast”● We want valleys “on the fly”● We need to find valley-prone subnets

55

Testing Persistence

0 205 10 15

56

consecutive trials

Testing Persistence

0 205 10 15

VS

Trial A Trial B

57

Latency Ratio Difference Over Time

58

Latency Ratio = (hop replica RTT) / (default replica RTT)

Testing Persistence

0 205 10 15

VS

Window A Window B

59

Testing Persistence

0 205 10 15

VS

Window A Window C

60

Testing Persistence

0 205 10 15

VS

Window A Window C

61

15 hours

Latency Ratio Difference Over Time

62

Latency Ratio = (hop replica RTT) / (default replica RTT)

Latency Ratio Difference Over Time

63

Latency Ratio = (hop replica RTT) / (default replica RTT)

Latency Ratio Difference Over Time

64

Latency Ratio = (hop replica RTT) / (default replica RTT)

Latency Ratio Difference Over Time

65

Latency Ratio = (hop replica RTT) / (default replica RTT)

SURPRISE! The Internet is crazy!

66

Filter: at least one valley

{0,0,0,0,0,V,0,0,0,0,0,0,V}

{0,0,0,0,0,0,0,0,0,0,0,0,0}

{V,V,V,V,0,0,0,0,V,V,V,0,V}

Subnet A

Subnet B

Subnet C

67

Filter: at least one valley

{0,0,0,0,0,V,0,0,0,0,0,0,V}

{0,0,0,0,0,0,0,0,0,0,0,0,0}

{V,V,V,V,0,0,0,0,V,V,V,0,V}

Subnet A

Subnet B

Subnet C

68

Filter: at least one valley

Latency Ratio = (hop replica RTT) / (default replica RTT)

69

Filter: at least one valley

Latency Ratio = (hop replica RTT) / (default replica RTT)

very flat

70

Filter: at least one valley

Latency Ratio = (hop replica RTT) / (default replica RTT)

very flat Close to zero

71

72

73

Parameter Exploration

How deep are the valleys from

useful subnets?

74

Vthresh =

Vthresh

75

10.6

0.9

Latency Ratio

Replicas

1

A B C

76

Vthresh 10.6

0.9

Latency Ratio

1

Replicas

A B C

How often do valleys occur in

useful subnets?

77

Vfreq =

78

TRAINING WINDOW

TRIALS79

80

Vfreq = 2/5

81

Vfreq = 2/5

Valley-Prone Subnet

82

Vfreq = 2/5

Valley-Prone Subnet

83

Vfreq = 2/5

NOT Valley-Prone Subnet

Overview of Drongo: 1. Collect training window

84

Overview of Drongo: 1. Collect training window

2. Count the # of sufficiently deep valleys

85

Overview of Drongo: 1. Collect training window

2. Count the # of sufficiently deep valleys

3. Apply subnet assimilationa. Training window is already completeb. Both parameters met

86

System Wide Performance

87

System Wide Performance

88

System Wide Performance

89

better

System Wide Performance

90

better

System Wide Performance

91

better

System Wide Performance

92

better

System Wide Performance

93

better

System Wide Performance

94

better

Vfreq = 1.0

System Wide Performance

95

better

Vfreq = 1.0

Vthresh = 0.95

96

Switch Quality

Global Params

Per Prov. Params

GoogleAmazon

AlibabaCDNetworks

ChinaNetCtr

CubeCDN

Conclusion & Insights

● CDNs have a lot of room for improvement

97

Conclusion & Insights

● CDNs have a lot of room for improvement

● Clients can help

98

Conclusion & Insights

● CDNs have a lot of room for improvement

● Clients can help

● Low requirements

99

Conclusion & Insights

● CDNs have a lot of room for improvement

● Clients can help

● Low requirements

● Can provide 50% improvement

100

Questions?

101

# Clients Affected

bette

r

102

Per Provider Overall Performance

103

Performance of Drongo’s choices

104