Drongo. Speeding Up CDNs with Subnet Assimilation from the Client. CoNEXT 17 Incheon, South Korea CDN & Caching Session

Transcription

1 Drongo CoNEXT 17 Incheon, South Korea CDN & Caching Session Speeding Up CDNs with Subnet Assimilation from the Client Authors: Marc Anthony Warrior Uri Klarman Marcel Flores Aleksandar Kuzmanovic

2 Bird s Eye View 2 What is Drongo? Why we need Drongo Performance Analysis Thoughts & Conclusions Questions 2

3 What is Drongo? 3

4 What is Drongo? It s a bird! 4

5 What is Drongo? It s a bird! 5

6 What is Drongo? It s a bird! 6

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) 7

8 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

9 Why Latency? 9

10 Latency is time Why Latency? 10

11 Latency is time Why Latency? Latency is money Google (Marissa Mayer), Amazon (Greg Linden) Web 2.0 Summet, glinden.blogspot.com 11

12 Latency is time Why Latency? 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

13 Drongo helps you (the end user) lower your own latency! 13

14 Drongo s Effect on Latency Go le og Am a n zo Ali b a ab CD N et rk wo tr s Ch ina tc Ne N CD e b Cu 14

15 Drongo s Effect on Latency le og Go Am a n zo Ali b a ab CD N et rk wo tr s Ch ina tc Ne N CD e b Cu ONLY client-side changes 15

16 Example Scenario 16

17 17 Provider wants to serve client

18 18 Client is far

19 19 CDN = more replica locations

20 20 DNS Redirection Which replica serves the client?

21 21 Choose the closest server

22 22 Choose the closest server This choice is nontrivial!

23 23 Often Suboptimal Choices!

24 Maybe just a far LDNS... [Chen - SigComm 15; Huang - SigComm CCR 12; Alzoubi - WWW 13; Rula - SigComm 14 ] 24

25 Ordinary DNS Query DNS Query LDNS IP Somewhere in California 25

26 EDNS0 Client-Subnet extension (ECS) DNS Query Somewhere in California LDNS IP Client Subnet Actually somewhere in New York 26

27 We used ECS: (ECS User) 27

28 We used ECS: This still happens (ECS User) 28

29 29 We used ECS: This still happens frequently (ECS User)

30 Really?... 30

31 Really?... YES! We measured it! 31

32 How did we measure it? 32

33 How did we measure it? Find subnets directed to different replicas 33

34 Subnet Assimilation DNS Query LDNS IP Client Subnet 34

35 Subnet Assimilation DNS Query LDNS IP Client Subnet Other Subnet 35

36 How did we measure it? Find subnets directed to different replicas 36 Perform pings and downloads to each replica

37 How did we measure it? Find subnets directed to different replicas 37 Perform pings and downloads to each replica Identify which subnet resulted in the best replica

38 38 1. Get Default Choice (use client s own subnet for ECS)

39 39 2. Traceroute to default choice

40 40 3. Get Hop Subnet Choices (use hops subnets for ECS)

41 41 4. Measure Latencies

42 42 4. Measure Latencies Steps 1-4: a trial

43 Latency Ratio 43 Normalize to default choice s RTT

44 We re looking for this

45 Valley = better choice from hop subnet replica choice for subnet traceroute 100 ms RTT: client to replica 0 ms 45

46 Valley = better choice from hop subnet replica choice for subnet traceroute 100 ms RTT: client to replica 0 ms 46

47 PlanetLab Sees Valleys! 47

48 PlanetLab Sees Valleys! 48

49 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

50 50 5.

51 51 5. Use best subnet for ECS

52 52 5. Use best subnet for ECS Get best mapping!

53 Are Valleys Predictable? Trials are not fast 53

54 Are Valleys Predictable? Trials are not fast We want valleys on the fly 54

55 Are Valleys Predictable? Trials are not fast We want valleys on the fly We need to find valley-prone subnets 55

56 Testing Persistence consecutive trials

57 Testing Persistence VS 0 5 Trial A Trial B 57

58 Latency Ratio Difference Over Time Latency Ratio = (hop replica RTT) / (default replica RTT) 58

59 Testing Persistence VS 0 5 Window A Window B 59

60 Testing Persistence VS 0 5 Window A Window C 60

61 Testing Persistence VS Window A Window C 15 hours 61

62 Latency Ratio Difference Over Time Latency Ratio = (hop replica RTT) / (default replica RTT) 62

63 Latency Ratio Difference Over Time Latency Ratio = (hop replica RTT) / (default replica RTT) 63

64 Latency Ratio Difference Over Time Latency Ratio = (hop replica RTT) / (default replica RTT) 64

65 Latency Ratio Difference Over Time Latency Ratio = (hop replica RTT) / (default replica RTT) SURPR ISE! The Internet is crazy! 65

66 Filter: at least one valley Subnet A {0,0,0,0,0,V,0,0,0,0,0,0,V} Subnet B {0,0,0,0,0,0,0,0,0,0,0,0,0} Subnet C {V,V,V,V,0,0,0,0,V,V,V,0,V} 66

67 Filter: at least one valley Subnet A {0,0,0,0,0,V,0,0,0,0,0,0,V} Subnet B {0,0,0,0,0,0,0,0,0,0,0,0,0} Subnet C {V,V,V,V,0,0,0,0,V,V,V,0,V} 67

68 Filter: at least one valley Latency Ratio = (hop replica RTT) / (default replica RTT) 68

69 Filter: at least one valley Latency Ratio = (hop replica RTT) / (default replica RTT) very flat 69

70 Filter: at least one valley Latency Ratio = (hop replica RTT) / (default replica RTT) very flat Close to zero 70

71 71

72 72

73 Parameter Exploration 73

74 How deep are the Vthresh = valleys from useful subnets? 74

75 Latency Ratio 1 Vthresh 0.6 A B C Replicas 75

76 Latency Ratio 1 Vthresh 0.6 A B C Replicas 76

77 How often do valleys Vfreq = occur in useful subnets? 77

78 TRAINING WINDOW 78

79 TRIALS 79

80 Vfreq = 2/5 80

81 Vfreq = 2/5 Valley-Prone Subnet 81

82 Vfreq = 2/5 Valley-Prone Subnet 82

83 Vfreq = 2/5 NOT Valley-Prone Subnet 83

84 Overview of Drongo: 1. Collect training window 84

85 Overview of Drongo: 1. Collect training window 2. Count the # of sufficiently deep valleys 85

86 Overview of Drongo: 1. Collect training window 2. Count the # of sufficiently deep valleys 3. Apply subnet assimilation a. Training window is already complete b. Both parameters met 86

87 System Wide Performance 87

88 System Wide Performance 88

89 System Wide Performance better 89

90 System Wide Performance better 90

91 System Wide Performance better 91

92 System Wide Performance better 92

93 System Wide Performance better 93

94 System Wide Performance Vfreq = 1.0 better 94

95 System Wide Performance Vfreq = 1.0 Vthresh = 0.95 better 95

96 Switch Quality G gle o o on A z ma ba Ali ba t C e DN tr C N t Ne CD a e b in Cu Ch rk wo s Global Params Per Prov. Params 96

97 Conclusion & Insights CDNs have a lot of room for improvement 97

98 Conclusion & Insights CDNs have a lot of room for improvement Clients can help 98

99 Conclusion & Insights CDNs have a lot of room for improvement Clients can help Low requirements 99

100 Conclusion & Insights CDNs have a lot of room for improvement Clients can help Low requirements Can provide 50% improvement 100

101 Questions? 101

102 better # Clients Affected 102

103 Per Provider Overall Performance 103

104 Performance of Drongo s choices 104