Kunal Shah Advisor: Dr. Harish Sethu

1. SIMULATION BASED STUDY OF TCP FAIRNESS IN MULTI-HOP WIRELESS NETWORKS Kunal Shah Advisor: Dr. Harish Sethu Computer Communications Laboratory

2. Outline • Introduction and Motivation • Background • Model Description • Simulation Results and Analysis • Conclusion

3. Introduction • Ad Hoc Networks: • A collection of nodes • Capable of acting as a host and a router simultaneously • Communicate with each other over shared, multi-hop wireless channels • Required where a fixed wired or wireless infrastructure is either unavailable or destroyed • Characterized by high mobility, low bandwidths, limited physical security and continuously changing network topology • Once the ad hoc network is up and running using some routing protocol, next step is to evaluate performance of transport layer protocol

4. Motivation • As local area wireless networks based on IEEE 802.11 standard see increasing public deployment, it is important to ensure that access to network by different users remains fair • No structured studies devoted to formal investigation of TCP fairness in wireless multi-hop networks

5. Focus • Evaluate TCP Tahoe, Reno, New Reno and SACK for fairness • Motivation for selecting these TCP implementations was their popularity • Fairness metric based on maximal normalized distance between user’s ideal share and actual share of service delivered by network • Analyze the effects of packet size, load, TCP receive buffer size and RTS/CTS on TCP fairness

6. TCP Evolution • Dominant reliable transport protocol since its origin • Consists of sliding window mechanism, which, in conjunction with ACKs and sequence numbers, guaranteed a reliable delivery and flow control • No congestion control or avoidance mechanism

7. TCP Evolution (Cont.) • AIMD – virtually the base of all existing TCP protocols • Besides maximizing link bandwidth, TCP must be fair to rest of the flows • Efficient TCP is not guaranteed to be fair

8. TCP Tahoe • Congestion Control Algorithms: • Slow Start • Congestion Avoidance • Fast Retransmit • Problem: • Transits to slow start after each packet loss

9. TCP Reno • Extension of TCP Tahoe • Added Fast Recovery along with Fast Retransmit TO: Time-out TD: Threshold Duplicate

10. TCP New Reno • TCP Reno problem: • Fast Recovery algorithm rendered inefficient in the presence of multiple losses within a single transmission window • TCP New Reno remains in fast recovery mode despite receiving partial acknowledgement after fast retransmission • Retransmits at the rate of one packet per RTT until all the lost packets are retransmitted • No retransmit timeout

11. TCP SACK • Selective Acknowledgements are used to provide the sender with sufficient information to recover from multiple packet losses within a single transmission window • Sender knows exactly which packets to retransmit and so is able to quickly recover • Problem: • Inefficient in the case of small sender window size

12. Fairness Criteria • Intuitively, one can think of fairness as the closeness of achieved throughput to its fair share

13. Max-Min Fairness (MMF) • When flows have equal weights, Max-Min Fair share allocation can be defined as: • Resources are allocated in order of increasing demand • No user gets a resource larger than its demand • Users with unmet demands get an equal share of the resource

14. MMF Example • Dividing a 8 slice pizza among 4 people 2 slices 2 slices 2 slices 4 slices 2 slices 2 slices + 1 slice = 3 slices desires and gets 3 slices 2 slices 2 slices + 1 slice = 3 slices 1 slice 4 slices 2 slices 2 slices - 1 slice = 1 slice

15. Ai – Fi U = maxi Fi Proposed Unfairness Criterion where Fi = MMFi (C, d1, d2, …, dn)

16. Sample Unfairness Calculation i di Ai MMFi 2 Mbps 0.8 Mbps 0.6 Mbps C = 1.8 Mbps 3 Mbps 0.7 Mbps 0.6 Mbps U = 0.5 1 Mbps 0.3 Mbps 0.6 Mbps

17. Related work on TCP Fairness • When flows with different end-to-end propagation delays shared a link, the bandwidth allocation was far from being fair • Constant rate window increase algorithm • Increase-by-K policy • Congestion Avoidance with Normalized Interval of Time (CANIT) • Wireless links are characterized by long RTT and above schemes react by opening up the congestion window at a much higher rate • Increased probing harmful as slow 56k modem links and band-limited wireless links are themselves a bottleneck in the network • Performance degradation not only due to transmission errors and losses but also due to congestion at base station → Fast TCP • Flow that got head-start occupied large amount of bandwidth and starved the flows starting later on→ split buffer queues • Unfair packet dropping policy at Internet routers→ RED policy

18. Related work on TCP Performance • Explicit Congestion Notification (ECN) • Explicit Link Loss Notification • M-TCP • Split TCP • Snoop TCP

19. MAC Layer Fairness • IEEE 802.11 uses per-node queue with per node back-off • Head-of-line packet headed towards a receiver that is in high contention neighborhood can block other flow transmissions to lightly loaded neighbors • Node with many flows penalizes its flow unfairly • Flows that experience more contention will block more contending flows while transmitting • Implementing changes made to MAC layer are impractical given the wide deployment of wireless networks using IEEE 802.11 standard • Lot of research done in improving MAC layer fairness and TCP performance but no real effort made in studying TCP fairness

20. Model Description - Node

21. Application Process Model

22. TCP Process Model

23. AODV Background • Source-based routing protocol based on DSDV and DSR • Utilizes sequence number of DSDV and on-demand route discovery and maintenance mechanisms of DSR • Power Efficient • No flooding or periodic update messages

24. AODV Process Model

25. WLAN Process Model

26. Simulation Scenario

27. Simulation Setup • Every TCP connection is of type FTP and all flows start at the same time • In each scenario, if user 1 (node 1) wants to send x Mbps, then user 2 wants to send 2x Mbps, user 3 wants to send 3x Mbps and so forth • Mobility pattern is static • Battery power is infinity and transmitter power is 0.25 Watts

28. Simulation Setup (Cont.) • Packet sizes are varied from 128 bytes to 1,024 bytes but ACKs are kept at 40 bytes long • TCP receive buffer size is varied from 8,760 bytes to 131,072 bytes • Load is varied from 1.5 Mbps to 7.5 Mbps to simulate low, medium and high traffic loads • Load is varied from 1.5 Mbps to 7.5 Mbps but with RTS/CTS enabled for packet sizes larger than 255 packets

29. Simulation Setup (Cont.) • All other parameters were left unchanged as per IEEE 802.11b standard • Simulation was conducted for TCP Tahoe, Reno, New Reno and SACK

30. Simulation Results – TCP Receive Buffer Size

31. Simulation Results – TCP Receive Buffer Size

32. Simulation Results – Load with No RTS/CTS

33. Simulation Results – Load with RTS/CTS

34. Simulation Results – Packet Size

35. Conclusion • Using the maximal normalized distance between the actual allocation and the max-min fair share allocation as a fairness metric, TCP fairness was evaluated for TCP Tahoe, Reno, New Reno and SACK by varying TCP receive buffer size, load with and without RTS/CTS and packet size • TCP Tahoe was the least unfair protocol but suffered from low throughput

36. Tentative Conclusions • Fairness best when: • TCP receive buffer size is large • Load is high and No RTS/CTS is deployed • Load is low and RTS/CTS is deployed • Packet size is larger for large TCP buffer size

37. Future Work • Tentative conclusions need to be studied in much more depth to comprehend the complex behavior of TCP in wireless networks • Analyze suggested TCP improvements like Split TCP and ECN for fairness • Introduce mobility and then analyze TCP fairness

38. Acknowledgements • I am sincerely grateful to my advisor Dr. Harish Sethu for watching, directing and guiding me throughout each stage of this work • I am thankful to Dr. Constantine Katsinis and Dr. Kapil Dandekar for serving in my thesis committee • I thank all the members of Computer Communications Laboratory for their support and responsiveness

39. Questions?