Optimizing Cost and Performance for Concurrent Multipath Transferring using extended shim6

1. Ye Wang, Xuan Li, Dongtao Liu, Maoke Chen ICCT2006 Guilin, China Optimizing Cost and Performance for Concurrent Multipath Transferring using extended shim6

2. Content • Introduction • CMT, shim6, traffic optimization • Modeling, Analysis, and Algorithm • lower cost when performance satisfied • higher performance when cost restricted • cost-performance tradeoff (higher NPR) • Simulation Results • Conclusion & Future Work ICCT2006

3. ISP1 CERNET2 Remote Site ISP2 NSFCNET Site ISP3 … Introduction – CMT • Concurrent Multipath Transferring • Multimedia / P2P / Distributed Computing / FTP / … 2001:980:ffe:1::42 2001:250::4321/64 2001:da8::5678/64 2001:9ef::0246/64 ICCT2006

4. Introduction – CMT • Possible CMT • multiple TCP connections • not applicable for general applications • SCTP • far beyond deployment • little work for end sites’ demand • TCP with extended shim6 • incremental deployable • more applicable • simple way to achieve traffic optimization ICCT2006

5. Introduction – shim6 • Site Multihoming by IPv6 Intermediation • Separate Identifier (host ID) from Locator (IP address) Receiver B Sender A ULP ULP Src ULID(A) Dst ULID(B) Src ULID(A) Dst ULID(B) shim6 Network Paths shim6 Src L(A) Dst L(B) Src L(A) Dst L(B) IP IP ICCT2006

6. Introduction – traffic optimization • Communication Solicitation of End Sites • Performance Maximization + Cost Minimization • Big Challenging for Internet • more than “best effort” • transferring in Layer 4, forwarding in Layer 3 • Our proposal: TCP CMT using extended shim6 • intermediation in Layer 4.5 • traffic (re)distribution on multiple Locator pairs • shim6 + TCP trigger ICCT2006

7. Introduction – traffic optimization • Related Works • Augmenting TCP • TCP-Reno, AQM/ECN, RED, XCP, VCP, etc. • SCTP: Stream Control Transmission Protocol • much more powerful than TCP • aim at better QoS for general purpose transferring • Smart routing • distribute traffic “optimally” in Layer 3 • long-term mechanism • gateway-side traffic control ICCT2006

8. Modeling – Definitions • Path i (i: 1~n): locator pair i • Traffic • traffic volume through Path i: • ISP Pricing and Site Cost • *flow pricing, lease pricing, flat-rate pricing • pricing via Path i by according ISP: • site cost: • Site Performance • bandwidth: • Equivalent Bandwidth: • Natural Price Ratio • NPR: 6 paths (n=6) Internet ISP21 ISP11 ISP12 ISP22 ISP13 ICCT2006

9. Modeling • Problem Statement • For what and not for what • How to (re)distribute TCP cwnd? care • How to rebuild TCP recv window? care • Which locator pairs to use? careless, depends on shim6 • Deal with fast retransmission? careless at present, depends on TCP-sub • 1. Optimize cost when performance satisfied • 2. Optimize performance when cost restrained • 3. Higher NPR ICCT2006

10. Analysis Typical demands for “at-least” bandwidth or “at-most” cost A simple metric of how well the traffic is distributed Cost NPR “Cost max” EB Cost “optimal point” “EB min” Goal: enhance performance and minimize cost ICCT2006

11. Analysis • Lemma • Timing pricing and lease pricing both call for performance maximization • Assumption • Paths are ordered by price (p1<p2<…<pn) • Theorem • N locator pairs, p1<p2<…<pn, max{ u } = max{ M / [ max{ mi/bi } * sum( pi*mi ) ] }  min max { mi/bi }  mi/bi = const = t b = b1 (1 + m2/m1 + m3/m1 + … + mn/m1)If mk>0, b = b1 + b2 + … + b(k-1) + b1*mk/m1( good for split and rebuild window, good performance )or min{ sum( pi*mi ) }m1 = m …( bad performance )or heuristic algorithm ICCT2006

12. Algorithms – 1 • Lower cost when performance satisfied • b(i)  m(i) / t (i = 1..N); • B  Bmin; i  1; • if B > b(i), then m(i)  b(i) * t; else m(i)  B * t, end; • B  B – b(i), i  i + 1 • if i <= n goto 3); else end. ICCT2006

13. Algorithms – 2 • Higher performance when cost restricted • Heuristic • reduce throughputs when deficit • improve performance when surplus • b(i)  m(i) / t (i = 1..n); • let C  sum( p(i) * m(i) ) (i = 1..n); • Cdelta  Cmax – C; • if Cdelta < 0 then m(i-1)  m(i) * 105% (i = 2..n); else m(i+1)  m(i) * 95% (i=1..n-1); ICCT2006

14. Algorithms – 3 • Higher NPR • similar as Algorithm 2, but use another metric for judgment • cut half load of Path i when • enlarge sub-window of Path i when • Relative Agio: • Quick metric: ICCT2006

15. ISP1 CERNET2 Remote Site ISP2 NSFCNET Site ISP3 … Simulation Results • Simulation in ns-2 • modified TCP agent + multiple modified IP agents • network topology: 3-ISP multihoming (+ N paths) • over 1000 runs ICCT2006

16. Simulation Results • High-bandwidth demand for CMT ICCT2006

17. Simulation Results • Constant-bandwidth demand for CMT ICCT2006

18. Simulation Results • Optimize cost when bandwidth satisfied ICCT2006

19. Simulation Results • Optimize bandwidth when cost restrained ICCT2006

20. Simulation Results • Optimize Natural Price Ratio ICCT2006

21. Conclusion and Future Work • Contribution • A case study on future IPv6 Internet technology (shim6) • Optimize CMT for the cost and performance of multihomed end sites within the TCP layer • Status • Good attempt in theory and simulation-based study • Little practice (shim6 has not yet been deployed), Only a best-effort way • Stepping forward • Develop real-system (shim6 prototype + UDP + algorithms) • Better congestion control mechanism (retransmission and reordering) • Study on site-ISP games ICCT2006

22. Ye Wang, Xuan Li, Dongtao Liu, Maoke Chen Thank you!