Performance-oriented Congestion Control

Size: px

Start display at page:

Download "Performance-oriented Congestion Control"

Adelia Randall
6 years ago
Views:

1 Performance-oriented Congestion Control Mo Dong, Qingxi Li, Doron Zarchy*, P. Brighten Godfrey and Michael Schapira* University of Illinois at Urbana Champaign *Hebrew University of Jerusalem

2 Mo Dong Qingxi Li Brighten Godfrey Doron Zarchy and Michael Schapira University of Illinois at Urbana-Champaign Hebrew University of Jerusalem

3 TCP Est

4 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP 3

5 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP 3

6 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP Who can be not happy with TCP? 3

7 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP are Who can be not happy with TCP? 3

8 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP 3

9 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP Point Solutions 4

10 TCP Est High BDP Wireless Satellite Inter-DC Intra-DC BIC H-TCP Compound CUBIC FAST TCP Westwood Vegas Veno Hybla STAR Illinois SABUL ICTCP DCTCP 10X 10X 17X 4X Unstable,RTT Unfair, Bufferbloat, Crash on Changing Networks,.. Point Solutions + Performance Far from Optimal 4

12 TCP fails to achieve consistent high performance

13 Why is it so hard?

14 6

15 Reno 1 pkt loss cwnd/2 Scalable ACK cwnd+1 CUBIC Time pass 1ms cwnd+f(t,cwn,rtt) FAST RTT increase x% Reduce cwnd to f(x)% HTCP 100 ACK cwnd+f(cwnd)/cwnd 6

16 Event Action Reno 1 pkt loss cwnd/2 Scalable ACK cwnd+1 CUBIC Time pass 1ms cwnd+f(t,cwn,rtt) FAST RTT increase x% Reduce cwnd to f(x)% HTCP 100 ACK cwnd+f(cwnd)/cwnd 6

17 Hardwired Mapping Event Action Reno 1 pkt loss cwnd/2 Scalable ACK cwnd+1 CUBIC Time pass 1ms cwnd+f(t,cwn,rtt) FAST RTT increase x% Reduce cwnd to f(x)% HTCP 100 ACK cwnd+f(cwnd)/cwnd 6

18 Hardwired Mapping Event Action 6

19 Event Action 7

20 Event Action Action 1 Network Action 2 N 7

21 Event Action Action 3 Network Action 4 7

22 Event Action Action 3 Network Action 4 7

23 Why? Event Action Action 3 Network Action 4 7

24 Why? Event Action Assumed Network Action 3 Network Action 4 7

25 Why? Event Action Assumed Network = Action 3 Real Network Network Action 4 Performance 7

26 Why? Event Action Assumed Network = Action 3 Real Network Network Action 4 Performance 7

27 Flow f sends at R Event Action 8

28 Flow f sends at R Event Action Packet Loss 8

29 Flow f sends at R Event Action Network Packet Loss 8

30 Flow f sends at R Event Action Network f causes most congestion Packet Loss 8

31 Flow f sends at R Event Action Network Dec R a lot f causes most congestion Packet Loss 8

32 Flow f sends at R Event Action Network Dec R a lot Packet Loss shallow buffer overflow 8

33 Flow f sends at R Event Action Network Dec R a lot Packet Loss shallow buffer overflow 8

34 Flow f sends at R Event Action Network Dec R a lot Packet Loss Dec R a little shallow buffer overflow 8

35 Flow f sends at R Event Action Network Dec R a lot Dec R a little Packet Loss other high rate flow causing congestion 8

36 Flow f sends at R Event Action Network Dec R a lot Dec R a little Packet Loss other high rate flow causing congestion 8

37 Flow f sends at R Event Action Network Dec R a lot Dec R a little Packet Loss Maintain R other high rate flow causing congestion 8

38 Flow f sends at R Event Action Network Dec R a lot Packet Loss Dec R a little Maintain R loss is random 8

39 Flow f sends at R Event Action Network Dec R a lot Packet Loss Dec R a little Maintain R loss is random 8

40 Flow f sends at R Event Action Network Dec R a lot Packet Loss Dec R a little Maintain R Increase R loss is random 8

41 Flow f sends at R Event Action Network Dec R a lot Packet Loss Dec R a little Maintain R Increase R loss is random 8

42 Flow f sends at R Event Action Network Dec R a lot Packet Loss Dec R a little Maintain R Increase R 8

43 Flow f sends at R Event Action Network Dec R a lot No event-control mapping optimal Packet Loss Dec R a little for all network scenarios Maintain R Increase R 8

44 What is the right rate to send? 9

45 What is the right rate to send? 9

46 What is the right rate to send? 9

47 What is the right rate to send? 9

48 What is the right rate to send? 9

49 What is the right rate to send? 10

50 What is the right rate to send? rate result 11

51 What is the right rate to send? rate r utility u U= f(tpt, loss rate, latency, etc.) e.g. U = tpt x (1 loss rate) 12

52 What is the right rate to send? rate r1 utility u1 U= f(tpt, loss rate, latency, etc.) e.g. U = tpt x (1 loss rate) No matter how complex the network, rate r > utility u 13

53 14

54 r1 14

55 r1 u1 14

56 r1 u1 r2 14

57 r1 u1 r2 u2 14

58 r1 r2 u1 u2 u1>u2? 14

59 r1 r2 u1 u2 u1>u2? move to r1 14

60 r1 r2 u1 u2 u1>u2? move to r1 move to r2 14

61 Performance-oriented Congestion Control move r1 r2 u1 u2 u1>u2? to r1 move to r2 15

62 Performance-oriented Congestion Control move r1 r2 u1 u2 u1>u2? to r1 move to r2 Observe real performance 15

63 Performance-oriented Congestion Control move r1 r2 u1 u2 u1>u2? to r1 move to r2 Observe real performance Control based on empirical evidence 15

64 Performance-oriented Congestion Control move r1 r2 u1 u2 u1>u2? to r1 move to r2 yields Observe real performance Control based on empirical evidence Consistent high performance 15

65 98 Mbps 102 Mbps this flow causing congestion X X X X X u1 u2 u1>u2? move to 98 Mbps 98 Mbps 102 Mbps random loss X X X u1 u2 u1>u2? move to 102 Mbps 16

66 Consistent High Performance 4X 17X 10X InterDC Throughput (Mbps) PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks 0.1 Throughput (Mbps) 100 PCC TCP illinois TCP CUBIC 10 Lossy Networks Relative throughput of long-rtt flow 1.2 Solves RTT Unfairness PCC TCP CUBIC TCP New Reno RTT of Long-RTT flow (ms) Throughput (Mbps) Bottleneck Buffersize (KB) 15X Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) Loss Rate Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Close to Optimal Time(s) Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks 17 5X in median Global Commercial Internet

67 Elephant 18

68 Where is Congestion Control? r1 r2 u1 u2 u1>u2? move to r1 move to r2 19

69 Where is Congestion Control? r1 r2 u1 u2 u1>u2? move to r1 move to r2 Selfishly maximizing utility 19

70 Where is Congestion Control? r1 r2 u1 u2 u1>u2? move to r1 move to r2 Selfishly maximizing utility => non-cooperative game 19

71 Where is Congestion Control? r1 r2 u1 u2 u1>u2? move to r1 move to r2 Selfishly maximizing utility => non-cooperative game Do we converge to a fair Nash equilibrium? 19

72 A class of utility functions converge to a fair and efficient Nash Equilibrium 20

73 A class of utility functions converge to a fair and efficient Nash Equilibrium u i (x) = T i * sigmoid(l i 0.05) x i * L i 20

74 A class of utility functions converge to a fair and efficient Nash Equilibrium u i (x) = T i * sigmoid(l i 0.05) x i * L i throughput sending rate observed loss rate 20

75 A class of utility functions converge to a fair and efficient Nash Equilibrium u i (x) = T i * sigmoid(l i 0.05) x i * L i Cut off loss rate at 5% Loss Rate 20

76 TCP Dynamics 21

77 TCP Dynamics AIMD as a hack to asymptotic fairness 21

78 TCP Dynamics AIMD as a hack to asymptotic fairness c 21 t

79 TCP Dynamics AIMD as a hack to asymptotic fairness c 21 t

80 TCP Dynamics AIMD as a hack to asymptotic fairness c 21 t

81 TCP Dynamics AIMD as a hack to asymptotic fairness c Moving away from convergence 21 t

82 PCC Dynamics c 22 t

83 PCC Dynamics PCC does not need AIMD because it looks at real performance c 22 t

84 PCC Dynamics PCC does not need AIMD because it looks at real performance c 22 t

85 PCC Dynamics PCC does not need AIMD because it looks at real performance c High Utility 22 t

86 PCC Dynamics PCC does not need AIMD because it looks at real performance c Game Theory Force High Utility 22 t

87 Throughput (Mbps) Throughput (Mbps) TCP Flow1 Flow2 Flow3 Flow PCC Convergence Time (s) Flow1 Flow2 Flow3 Flow Time (s) 23

88 Consistent High Performance 24

89 Consistent High Performance 100 Throughput (Mbps) 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Bottleneck Buffersize (KB) 24

90 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) 24

91 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) Loss Rate PCC TCP illinois TCP CUBIC 10 Lossy Networks 24

92 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate 24

93 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Throughput (Mbps) Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) 24

94 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) 24

95 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) 1000 Goodput (Mbps) IntraDC Incast PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb Number of Senders 24

96 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) Number of Senders 24

97 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) 24

98 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Buffer Size(KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) 24

99 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Relative throughput of long-rtt flow Buffer Size(KB) RTT Unfairness PCC TCP CUBIC TCP New Reno RTT of Long-RTT flow (ms) 24

100 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Relative throughput of long-rtt flow Buffer Size(KB) 1.2 Solves RTT Unfairness PCC TCP CUBIC TCP New Reno RTT of Long-RTT flow (ms) 24

101 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Throughput (Mbps) 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Relative throughput of long-rtt flow Buffer Size(KB) 1.2 Solves RTT Unfairness PCC TCP CUBIC TCP New Reno RTT of Long-RTT flow (ms) InterDC 24

102 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Throughput (Mbps) 4X vs Illinois 1.23X vs UDT InterDC 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Relative throughput of long-rtt flow Buffer Size(KB) 1.2 Solves RTT Unfairness PCC TCP CUBIC TCP New Reno RTT of Long-RTT flow (ms) 24

103 0.1 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Throughput (Mbps) 4X vs Illinois 1.23X vs UDT InterDC 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Relative throughput of long-rtt flow Buffer Size(KB) 1.2 Solves RTT Unfairness PCC TCP CUBIC TCP New Reno RTT of Long-RTT flow (ms) Global Commercial Internet 24

104 0.1 Consistent High Performance X vs Hybla 10 PCC TCP Hybla TCP Illinois 1 TCP CUBIC TCP New Reno Satellite Networks Throughput (Mbps) Bottleneck Buffersize (KB) Similar to ICTCP Goodput (Mbps) IntraDC Incast Number of Senders PCC data=256kb PCC data=128kb PCC data=64kb TCP data=256kb TCP data=128kb TCP data=64kb SendingRate(Mbps) Throughput (Mbps) 4X vs Illinois 1.23X vs UDT InterDC 100 PCC TCP illinois 10X vs Illinois TCP CUBIC 10 Lossy Networks Loss Rate Close to Optimal PCC TCP Illinois TCP CUBIC Rapidly Changing Networks Time(s) Throughput (Mbps) 15X less buffer Shallow Network Buffer 20 PCC 10 TCP Pacing 0 TCP CUBIC Solves RTT Unfairness PCC TCP CUBIC TCP New Reno X median vs CUBIC 1.48X vs UDT + 4X less loss Global Commercial Internet Relative throughput of long-rtt flow Buffer Size(KB) RTT of Long-RTT flow (ms) 24

105 Consistent High Performance Global Commercial Internet 25

106 PCC vs TCP vs Take a Flight 100G Data 26

107 PCC vs TCP vs Take a Flight Utah, U.S. > Berlin, Germany Illinois, US > Waseda, Japan Georgia, US > Stockholm, Sweden Georgia, US > Ljubljana, Slovenia Missouri, US > Rennes, France Massachusetts US > Seoul, Korea 100G Data PCC TCP CUBIC Take a Flight 0:0:00:00 1:18:27:30 3:12:55:00 5:7:22:30 7:1:50:00 26

108 PCC vs TCP vs Take a Flight Utah, U.S. > Berlin, Germany Illinois, US > Waseda, Japan Georgia, US > Stockholm, Sweden Georgia, US > Ljubljana, Slovenia Missouri, US > Rennes, France Massachusetts US > Seoul, Korea 100G Data PCC TCP CUBIC Take a Flight 0:0:00:00 1:18:27:30 3:12:55:00 5:7:22:30 7:1:50:00 26

109 PCC vs TCP vs Take a Flight Utah, U.S. > Berlin, Germany Illinois, US > Waseda, Japan Georgia, US > Stockholm, Sweden Georgia, US > Ljubljana, Slovenia Missouri, US > Rennes, France Massachusetts US > Seoul, Korea 100G Data PCC TCP CUBIC Take a Flight 0:0:00:00 1:18:27:30 3:12:55:00 5:7:22:30 7:1:50:00 26

110 PCC vs TCP vs Take a Flight Utah, U.S. > Berlin, Germany Illinois, US > Waseda, Japan Georgia, US > Stockholm, Sweden Georgia, US > Ljubljana, Slovenia Missouri, US > Rennes, France Massachusetts US > Seoul, Korea 100G Data PCC TCP CUBIC Take a Flight 0:0:00:00 1:18:27:30 3:12:55:00 5:7:22:30 7:1:50:00 26

111 Consistent High Performance Rapidly Changing Networks BW: Mbps; RTT: ms; Loss Rate: 0-1% Change every 5 seconds Sending Rate (Mbps) Time(s) Optimal 27

112 Consistent High Performance Sending Rate (Mbps) Rapidly Changing Networks BW: Mbps; RTT: ms; Loss Rate: 0-1% Change every 5 seconds Time(s) Optimal TCP Illinois TCP CUBIC 28

113 Consistent High Performance Rapidly Changing Networks BW: Mbps; RTT: ms; Loss Rate: 0-1% Change every 5 seconds SendingRate(Mbps) Time(s) Optimal PCC TCP Illinois TCP CUBIC 29

114 Long list of things we have done but don t have time to talk about More stories about the fact that TCP is broken Proof of Nash Equilibrium and Convergence Concrete Implementation of PCC Performance monitoring Details of learning control algorithm Implementation designs and optimizations Performance Evaluation Plaenetlab detailed explanation Satellite like network Lossy Links TCP friendliness RTT fairness Shallow buffer networks Inter data center networks small buffer networks Reactiveness and stability tradeoff Jain index fairness Benefit of Randomized Control Trials Details of TCP friendliness evaluation Emulated satellite networks Emulated datacenter networks Cure RTT unfairness Does not fundamentally harm short flow FCT Evaluation in the wild vs non-tcp protocols Flexibility by pluggable utility function 30

115 Cellular Networks Deep queue and bufferbloat Highly dynamic available bandwidth PCC Flexible (latency sensitive) utility function Handling small causality window 31

116 Flexible Utility Function 32

117 Flexible Utility Function Same Rate Control Algorithm 32

118 Flexible Utility Function Same Rate Control Algorithm + 32

119 Flexible Utility Function Same Rate Control Algorithm + Different Utility Function = Flexibility 32

120 Latency Sensitive Utility Function Power = Tpt/RTT TCP Latency Sensitive Deep Queue 33

121 Latency Sensitive Utility Function Power = Tpt/RTT TCP Latency Sensitive Deep Queue Codel 33

122 Latency Sensitive Utility Function Power = Tpt/RTT TCP Latency Sensitive Deep Queue Codel 33

123 Latency Sensitive Utility Function Power = Tpt/RTT TCP PCC Latency Sensitive Deep Queue Codel 33

124 Small Causality Window r1 r2 u1 u2 u1>u2? move to r1 move to r2 34

125 Small Causality Window Figure drawn using Mahimahi by K. Winstein et. al. ( 35

126 Small Causality Window PCC V2 Fast convergence to optimal in tiny causality window Fast reaction to causality disruption Handling noisy measurement 36

127 Rate control based on empirically observed performance yields Consistent high performance Better stability than TCP Flexible performance objectives PCC V2 is coming 37

128 38

129 speedier.net 38

130 speedier.net/pcc 39

131 40

132 Backup Slides 40

133 Software Components Sender Performance Oriented Control Module Sending Rate Control Sending Module Data Receiver (Sending Rate, Utility) Sent Packet Log Network Utility Function Performance Metrics (tpt., loss rate, RTT) Monitor Module SACK 41

134 Software Components Sender Performance Oriented Control Module Sending Rate Control Sending Module Data Receiver (Sending Rate, Utility) Sent Packet Log Network Utility Function Performance Metrics (tpt., loss rate, RTT) Monitor Module SACK 41

135 Performance Oriented Control observed utility rate 42

136 Performance Oriented Control observed utility rate 42

137 Performance Oriented Control observed utility rate 42

138 Performance Oriented Control observed utility rate 43

139 Performance Oriented Control observed utility? r rate 43

140 Performance Oriented Control observed utility? r(1 ε) r rate 43

141 Performance Oriented Control observed utility? r(1 ε) r r(1+ ε) rate 43

142 Performance Oriented Control observed utility? randomized controlled trials r(1 ε) r r(1+ ε) rate 43

143 Performance Oriented Control observed utility? randomized controlled trials r(1 ε) r r(1+ ε) rate 43

144 Performance Oriented Control observed utility? randomized controlled trials r(1 ε) r r(1+ ε) rate 43

145 Performance Oriented Control observed? utility?? randomized controlled trials r(1 ε) r r(1+ ε) rate 43

146 Performance Oriented Control observed? utility?? randomized controlled trials ε min = 0.01 r r(1 2ε) r(1+ 2ε) rate 43

147 Online Learning Control observed utility r 0 rate 44

148 Online Learning Control observed utility r 0 r 0 (1+ 2ε) r 1 rate 44

149 Online Learning Control observed utility r 0 r 0 (1+ 2ε) r 1 (1+ 3ε) r 1 r 2 rate 44

150 Online Learning Control observed utility r 0 r 0 (1+ 2ε) r 1 (1+ 3ε) r 2 (1+ 4ε) r 1 r 2 r 3 rate 44

151 Online Learning Control observed utility? r 0 r 0 (1+ 2ε) r 1 (1+ 3ε) r 2 (1+ 4ε) r 1 r 2 r 3 rate 44

152 Deployment No hardwired support, packet header, protocol change needed Where to deploy CDN backbone, Inter-data center, dedicated scientific nw In the Wild? 45

153 TCP Friendliness 46

154 TCP Friendliness Is PCC TCP-friendly? 46

155 TCP Friendliness Is PCC TCP-friendly? Wrong Question 46

156 TCP Friendliness 47

157 TCP Friendliness PCC s default utility function is not TCP Friendly 47

158 TCP Friendliness PCC s default utility function is not TCP Friendly But not that bad 47

159 TCP Friendliness PCC s default utility function is not TCP Friendly But not that bad 47

160 TCP Friendliness PCC s default utility function is not TCP Friendly But not that bad IE11 47

161 TCP Friendliness PCC s default utility function is not TCP Friendly Different utility functions can be a solution 48

162 TCP Friendliness PCC s default utility function is not TCP Friendly Different utility functions can be a solution u i (x) = T i * sigmoid α (L i 0.05)* sigmoid β ( rtt n rtt n 1 1) x i * L i rtt n 48

163 TCP Friendliness PCC s default utility function is not TCP Friendly Different utility functions can be a solution 49

164 TCP Friendliness PCC s default utility function is not TCP Friendly Different utility functions can be a solution TCP vs TCP TCP vs PCC 49

165 TCP Friendliness PCC s default utility function is not TCP Friendly Different utility functions can be a solution TCP vs TCP TCP vs PCC 49

166 TCP Friendliness PCC s default utility function is not TCP Friendly Different utility functions can be a solution TCP vs TCP TCP vs PCC 49

167 Deployability: Short Flow FCT PCC does not fundamentally harm FCT Flow Completion Time (ms) PCC-95th TCP-95th PCC-average TCP-average PCC-median TCP-median Network Load (%)

168 Consistent High Performance 100 Global Commercial Internet Percentage of Trials Throughput Improvement Ratio 51 PCC vs TCP CUBIC

169 Consistent High Performance 100 Global Commercial Internet Percentage of Trials %, 10X Throughput Improvement Ratio 51 PCC vs TCP CUBIC

170 Consistent High Performance Percentage of Trials Global Commercial Internet Throughput Ratio over TCP CUBIC PCC_d 52

171 Consistent High Performance Percentage of Trials Global Commercial Internet Default utility function Throughput Ratio over TCP CUBIC PCC_d 52

172 Consistent High Performance Percentage of Trials Global Commercial Internet Default utility function Throughput Ratio over TCP CUBIC PCC_f PCC_d 53

173 Consistent High Performance Percentage of Trials Global Commercial Internet Friendly Utility Function Default utility function Throughput Ratio over TCP CUBIC PCC_f PCC_d 53

174 Consistent High Performance Global Commercial Internet Friendly Utility Function Default utility function 54

175 Intra-continent 2.33X median Consistent High Performance Global Commercial Internet Friendly Utility Function Default utility function Inter-continent 25X median 54

176 Consistent High Performance Global Commercial Internet 55

177 Consistent High Performance Global Commercial Internet 1.48X Median 4X less loss 55

178 Consistent High Performance Satellite Network 17X WINDS System 56

179 Consistent High Performance Lossy Networks 100Mbps, 30ms, varying loss rate TCP s throughput collapse 10X with 0.5% loss 19X 57

180 Consistent High Performance Treat RTT Unfairness 100Mbps, 10ms short RTT flow Varying RTT of long RTT one 58

181 Consistent High Performance A possible solution to the bufferbloat problem 100Mbps, 30ms Link 90 pkts, 90% tpt 6 pkts, 90% tpt 59

182 Consistent High Performance Inter Datacenter and Dedicated High Speed Network 60

183 Consistent High Performance Inter Datacenter and Dedicated High Speed Network 123% 60

184 Consistent High Performance Mitigate Incast 61

185 Long list of things we have done but don t have time to talk about 62

Performance-oriented Congestion Control

Performance-oriented Congestion Control Mo Dong, Qingxi Li, Doron Zarchy*, P. Brighten Godfrey and Michael Schapira* University of Illinois at Urbana Champaign *Hebrew University of Jerusalem Qingxi Li