1. Internet Protocols Steven Low CS/EE netlab.CALTECH.edu October 2004 with J. Doyle, L. Li, A. Tang, J. Wang

2. Internet Protocols Selects paths from sources to dests TCP/AQM IP x i (t) p l (t) Selects source transmission rates Link Selects topology, capacities, power,… Application Selects user criteria, web layout, utility

3. Internet Protocols TCP/AQM IP x i (t) p l (t) Link Application  Protocols determines network behavior  Critical, yet difficult, to understand and optimize  Local algorithms, distributed spatially and vertically  global behavior  Designed separately, deployed asynchronously, evolves independently

4. Internet Protocols TCP/AQM IP x i (t) p l (t) Link Application  Protocols determines network behavior  Critical, yet difficult, to understand and optimize  Local algorithms, distributed spatially and vertically  global behavior  Designed separately, deployed asynchronously, evolves independently Need to reverse engineer … to forward engineer new large networks

5. Internet Protocols TCP/AQM IP x i (t) p l (t) Link Application  Protocols determines network behavior  Critical, yet difficult, to understand and optimize  Local algorithms, distributed spatially and vertically  global behavior  Designed separately, deployed asynchronously, evolves independently Need to reverse engineer … much easier than biology with full specs

6. Internet Protocols Minimize path costs (IP) TCP/AQM IP x i (t) p l (t) Maximize utility (TCP/AQM) Link Minimize SIR, max capacities, … Application Minimize response time (web layout)

7. Internet Protocols  Each layer is abstracted as an optimization problem  Operation of a layer is a distributed solution  Results of one problem (layer) are parameters of others  Operate at different timescales TCP/AQM IP Link Application IPLink

8. IP TCP/ AQM Applications Link Chiang: X-ities?? Utility maximization Chiang: Wireless issues ?? Outline General Approach: 1) Understand a single layer in isolation and assume other layers are designed nearly optimally. 2) Understand interactions across layers 3) Incorporate additional layers, with the ultimate goal of viewing entire protocol stack as solving one giant optimization problem (where individual layers are solving parts of it).

9. IP TCP/ AQM Applications Link Outline Utility maximization Some implications

10. Network model F1F1 FNFN G1G1 GLGL R R T TCP Network AQM x y q p Reno, Vegas DT, RED, … IP routing

11. Network model - example Reno, Vegas DT, RED, … IP routing TCP Reno: currently deployed TCP AI MD TailDrop

12. Network model - example Reno, Vegas DT, RED, … IP routing TCP FAST: high speed version of Vegas

13. Duality model Flow control problem (Kelly, Malloo, Tan 98) TCP/AQM Maximize utility (& solve Dual) with different utility functions (L 03): (x*,p*) primal-dual optimal iff Primal-dual algorithm Reno, Vegas, FASTDT, RED, REM/PI, AVQ

14. Duality model Historically  Packet level implemented first  Flow level understood as after-thought  But flow level design determines performance, fairness, stability  Vegas, FAST, STCP  HSTCP (homogeneous sources)  Reno (homogeneous sources)  infinity  XCP (single link only) (Mo & Walrand 00)

15. Duality model Historically  Packet level implemented first  Flow level understood as after-thought  But flow level design determines performance, fairness, stability Now: can forward engineer Given (application) utility functions, can generate provably scalable TCP algorithms

16. Protocol decomposition TCP-AQM: TCP algorithms maximize utility with different utility functions Congestion prices coordinate across protocol layers TCP-AQM IP TCP/ AQM Applications Link

17. Protocol decomposition TCP-AQM IP TCP-AQM IP TCP/ AQM Applications Link TCP/IP: TCP algorithms maximize utility with different utility functions Shortest-path routing is optimal using congestion prices as link costs Congestion prices coordinate across protocol layers

18. Protocol decomposition TCP-AQM IP TCP/IP (with fixed c ): Equilibrium of TCP/IP exists iff zero duality gap NP-hard, but subclass with zero duality gap is LP Equilibrium, if exists, can be unstable Can stabilize, but with reduced utility Inevitable tradeoff bw utility max & routing stability TCP-AQM IP TCP/ AQM Applications Link

19. Protocol decomposition IP Link TCP/IP with optimal c : With optimal provisioning, static routing is optimal using provisioning cost  as link costs TCP/IP with static routing in well-designed network TCP-AQM IP TCP-AQM IP TCP/ AQM Applications Link

20. IP TCP/ AQM Applications Link Outline Utility maximization Some implications

21. Implications Is fair allocation always inefficient ? Does raising capacity always increase throughput ? Intricate and surprising interactions in large-scale networks … unlike at single link

22. Implications Is fair allocation always inefficient Does raising capacity always increase throughput

23. Fairness Identify allocation with  An allocation is fairer if its  is larger (Mo, Walrand 00)

24. Fairness  maximum throughput  proportional fairness  min delay fairness (Reno)  infinity  maxmin fairness (Mo, Walrand 00)

25. Efficiency Unique optimal rate x () An allocation is efficient if T () is large

26. Conjecture T () is nonincreasing i.e. a fair allocation is always inefficient

27. Example 1 Conjecture T () is nonincreasing i.e. a fair allocation is always inefficient 0 max throughput 1 1/(L+1) proportional fairness L/L(L+1) 1/2 maxmin fairness 1/2

28. Intuition “The fundamental conflict between achieving flow fairness and maximizing overall system throughput….. The basic issue is thus the trade-off between these two conflicting criteria.” Luo,etc.(2003), ACM MONET

29. Results  Theorem: Necessary & sufficient condition for general networks (R, c) provided every link has a 1-link flow  Corollary 1: true if N(R)=1

30. Counter-example  There exists a network such that dT/d > 0 for almost all >0  Intuition Large  favors expensive flows Long flows may not be expensive  Max-min may be more efficient than proportional fairness expensive long

31. Counter-example  Theorem: Given any  0 >0, there exists network where  Compact example

32. Implications Is fair allocation always inefficient ? Does raising capacity always increase throughput ? Intricate and surprising interactions in large-scale networks … unlike at single link

33. Throughput & capacity  Intuition: Increasing link capacities always raises throughput T  Theorem: Necessary & sufficient condition for general networks (R, c)  Corollary: For all , increasing a link’s capacity can reduce T all links’ capacities equally can reduce T all links’ capacities proportionally raises T

34. Protocol decomposition TCP-AQM IP Link TCP algorithms maximize utility with different utility functions IP shortest path routing is optimal using congestion prices as link costs, with given link capacities c With optimal provisioning, static routing is optimal using provisioning cost  as link costs Congestion prices coordinate across protocol layers

35. Some implications TCP-AQM TCP/AQM: TCP maximizes aggregate utility (not throughput) Fair bandwidth allocation is not always inefficient Increasing capacity does not always raise throughput Intricate network interactions  paradoxical behavior