Tweet
Tweet

Slide 1

Slide 1 text

Jingyuan Wang, Jiangtao Wen, Jun Zhang and Yuxing Han Tsinghua National Laboratory for Information Science and Technology

Slide 2

Slide 2 text

 Introduction  Background and Motivation  The TCP-FIT Algorithm  Experimental Results  Conclusion

Slide 3

Slide 3 text

 TCP ◦ The Transmission Control Protocol  TCP Reno/ TCP New Reno ◦ Pros  Reliable  Sequential ◦ Cons  Wireless network  high Bandwidth Delay Product (BDP)

Slide 4

Slide 4 text

 Recent development ◦ Wireless  TCP Westwood  TCP Veno ◦ BDP  TCP Compound  TCP CUBIC  FAST TCP  TCP-FIT ◦ Inspired by parallel TCP

Slide 5

Slide 5 text

 Introduction  Background and Motivation  The TCP-FIT Algorithm  Experimental Results  Conclusion

Slide 6

Slide 6 text

 What is a Congestion Algorithm?  Classification ◦ Loss-based ◦ Delay-based ◦ Hybrid

Slide 7

Slide 7 text

 Loss-Based ◦ TCP Reno, TCP CUBIC ◦ TCP BIC , High Speed TCP  Definition  Assumption ◦ Constraints  In wireless scenarios  In BDP scenarios

Slide 8

Slide 8 text

 Delay-Based ◦ TCP Vegas and FAST TCP  Definition  Assumption ◦ Queuing delay = RTT – Propagation delay. ◦ Pros  Resilient -> good ◦ Cons  bandwidth starvation

Slide 9

Slide 9 text

 Hybrid TCP ◦ Veno, TCP Westwood, TCP Illinois, H-TCP and Fusion TCP ◦ Compound TCP  Cons ◦ Poor in BDP-wireless-hybrid scenarios

Slide 10

Slide 10 text

 TCP-FIT ◦ Parallel TCP ◦ E.g. GridFTP and E-MulTCP ◦ Pros:  Utilization  Good in wireless and BDP ◦ Cons:  Compatibility  Fairness

Slide 11

Slide 11 text

 Introduction  Background and Motivation  The TCP-FIT Algorithm  Experimental Results  Conclusion

Slide 12

Slide 12 text

 Notation of parameter ◦ : Size of congestion window  � ◦ : RTT time ◦ : PLR, Packet Loss Rate ◦ : Throughput of the network ◦ : Propagation delay ◦ : Queuing delay

Slide 13

Slide 13 text

 Object ◦ achieve N times throughput of the TCP Reno ◦ meanwhile maintain fairness  TCP Reno ◦ AIMD  : ← + 1  : ← − 1 2

Slide 14

Slide 14 text

 MulTCP ◦ : ← + ◦ : ← − 1 2  Proposed Method ◦ : ← + ◦ : ← − 2 3+1

Slide 15

Slide 15 text

 Assumption: one loss adjustment is enough ◦ Based on AIMD ◦ = − 2 3+1 ◦ + = 2� ◦ We get ◦ = 3+1 3 �

Slide 16

Slide 16 text

 In a certain length of time( we set it as k*RTT time ) ◦ ∆ = 1 − � + � 𝑘𝑘 � � 2 3+1  Then ◦ ̇ = ∆ ̇ = 𝑇𝑇 − 𝑋𝑋 � 2 3+1 � 3+1 3 �  Let ̇ = 0 then we get = 32 2 � 2 ( = � 𝑇𝑇 ) ◦ = 1 𝑇𝑇 3 2 1

Slide 17

Slide 17 text

 The is adjusted adaptively. Update of is ◦ 𝑖𝑖+1 = 1, 𝑖𝑖 + − −𝑡𝑡 𝑖𝑖  The equilibrium of is ◦ = � −𝑡𝑡 = 𝑇𝑇 ◦ (denote α as α = (𝑡𝑡 − 𝑡𝑡 )/𝑡𝑡 ) ◦ (Easy to know = = + = 𝑡𝑡 + ) ◦ = 2 3 1

Slide 18

Slide 18 text

 Network Utilization ◦ = 1 𝑇𝑇 3 2 1 vs = 1 𝑇𝑇 3 2 1  Fairness ◦ RTT-fairness  η = 𝑖𝑖 = 1 vs η = 𝑖𝑖 = + 𝑖𝑖+ ( = = + ) ◦ Inter-fairness  = 𝑇𝑇∗−𝑇𝑇′ 𝑇𝑇∗

Slide 19

Slide 19 text

 Introduction  Background and Motivation  The TCP-FIT Algorithm  Experimental Results  Conclusion

Slide 20

Slide 20 text

 BDP Scenarios

Slide 21

Slide 21 text

 Wireless Scenarios

Slide 22

Slide 22 text

 Inter-fairness

Slide 23

Slide 23 text

 RTT-fairness

Slide 24

Slide 24 text

 Cons: ◦ a low speed ADSL network with large bandwidth variations ◦ Due to simplistic model of bandwidth estimation compared with FAST.

Slide 25

Slide 25 text

Thank you