1. Alaa M. Kharma 0041610 Abdelrahman N. El-Sharif 0020586 Special Topics in Computer Engineering Instructor: Dr. Walid Abu-Sufah

2. VIOLETA GAMBIROZA, BAHAREH SADEGHI, AND EDWARD W. KNIGHTLY DEPARTMENT OF ELECTRICAL AND COMPUTER ENGINEERING RICE UNIVERSITY HOUSTON, TX 77005 End to End Performance and Fairness in Multihop Wireless Backhaul Networks

3. Our Objective Is To study fairness and end-to-end performance in multihop wireless backhaul networks  Backhaul Networks: transmitting from a remote site or network to a central or main site. For example, from a wireless mesh network to the wired network means aggregating all the traffic on the wireless mesh over one or more high-speed lines to the Internet Internet

4. Why Use Wireless Backhaul Networks ? High bandwidth  High spatial reuse and capacity scaling  Opportunistic protocols High availability  Redundant paths and non-mobile infrastructure  Deployability Good economics  Unlicensed spectrum  Few wires

5. Multihop Wireless Backhaul Networks Problems l Existing protocols result in l Severe unfairness l Poor performance l Starvation for users located more than one hop away from the wired entry point l “Hidden terminals” and “information asymmetry” l Spatial bias

6. Methodology Developing a formal reference model that characterizes objectives such as removing spatial bias  Providing performance that is independent of the number of wireless hops to a wire  Maximizing spatial reuse Performing an extensive set of simulation experiments Developing and studying a distributed layer 2 fairness algorithm which targets to achieve the fairness of the reference model without modification to TCP. Studying the critical relationship between fairness and aggregate throughput  In particular study the fairness-constrained system capacity of multihop wireless backhaul networks.

7. Fairness Reference Model Used as a target and benchmark for protocol design and as a tool for studying alternatives for a network’s fairness and performance objectives

8. TAP Fairness Reference Model objectives TAP’s traffic should be treated as a single aggregate Maximal spatial reuse must be ensured when network resources are unused  Either due to lack of demand or in cases of sufficient demand in which flows are bottlenecked elsewhere Spatial bias must be eliminated to ensure that nodes one hop away from a wired entry point do not receive a disproportionately greater share of resources than nodes multiple hops away Time rather than throughput should be considered as the basic network resource that needs to be fairly shared

9. Key Performance Factors Fairness algorithms  [UDP, TCP, and IFA] Media access protocols  [CSMA and CSMA/CA] Channel models  [constant rate, ricean] Antenna technologies  [Omni directional, sector] Multiple topologies and flow scenarios  Parking Lot

10. Some Definitions TAP: Transit Access Points MU: Mobile Unit TA(i): The aggregate traffic flow ingressing at TAP Spatial Reuse: Transmission between the different pairs simultaneously if they do not cause no interference

11. Goodput The number of useful bits per unit of time forwarded by the network from a certain source address to a certain destination, excluding protocol overhead, and excluding retransmitted data packets. The goodput is generally lower than the throughput Factors that may cause lower goodput than throughput are:  Protocol overhead.  Retransmission of lost or corrupt packets  Collision detection in the Ethernet CSMA/CD protocol, and collision avoidance may cause "backoff" waiting time and retransmission

12. Parking Lot Scenario

13. Parking Lot Scenario Cont.

15. UDP Baseline Scenario Cont. Parking Lot Scenario

16. UDP Baseline Scenario Cont. The bars labeled “Obj.” depicts the shares specified by the reference model

17. Observations The starvation occurs because  TAP1 experiences exponential back-off far more frequently than TAP2 and TAP3 due to the “hidden terminal” problem  This results in collision at TAP2 and exponential back-off for TAP1 Flows originating at far hops contend for the channel an increasing number of times  Higher probability of collision and loss  Corresponding throughput decreases The system has achieved 92% of the aggregate capacity of the reference model Must consider capacity and fairness in performance analysis  High aggregate capacity in the presence of starved flows is clearly undesirable

18. Hidden Terminals Internet TAP1 TAP2 TAP3 TAP4 collisionno collision

19. TCP Fairness As TCP acknowledgment packets form a traffic aggregate in the reverse direction, we depict this traffic as TAP4’s goodput

20. Observations Unable to prevent traffic originating at TAP1 from starving Tap2 traffic is now starved as well due to effects of TCP The key reasons for this poor performance are  Both TAP1 and TAP2 are now hidden terminals since TCP acks generate traffic in the opposite direction  Losses can lead to timeouts, resulting in a congestion window of 1 segment and a significant throughput penalty

21. Observations Cont. CSMA/CA exchanges RTS/CTS results in  A decreased number of collisions, however introduces the problem of “information asymmetry” CSMA obtains slightly higher goodput as compared to CSMA/CA  As it does not incur overhead due to RTS/CTS exchange Total goodput for TCP traffic is higher than that of the reference model objective  Starving multihop flows and giving all capacity to one hop flows is indeed the capacity-maximizing allocation

22. Information Asymmetry Internet TAP1 TAP2 TAP3 TAP4 RTS TAP2 sets its NAV No CTS RTS Node with more information knows when to contend; other attempts randomly

23. Flow Scenario Parallel Parking Lot  Transmission between the pairs TAP1-TAP2 and TAP4 - TAP5 can occur simultaneously (spatial reuse)

24. Flow Scenario

25. Observations Flows (1,5), TA(2), and TA(3) are starved While flow (1,2) is indeed able to exploit spatial reuse, it has done so only because TAP3 traffic is starved  If this traffic was not starved, TAP3 would be a hidden terminal for TAP1 and would result in significant performance degradation

26. Sector Antennas Sector antennas provide statically-configured directional transmission and reception that results in  Increased spatial reuse  Increased transmission range  All sectors can be active simultaneously  Each sector has its own air interface and MAC

27. Sector Antennas

28. Observations The reference model obtains throughputs for each TAP double that of the prior case due to the second antenna and MAC Sector antennas have eliminated the hidden terminal problem and asymmetry problem Total goodput is increased as compared to the omnidirectional case  TAP1 and TAP2 are no longer starved and now contribute to goodput  There is a second air interface  There are reduced collisions

29. Channel Model Multi-rate Transmission  Adapt their transmission rate according to SNR Taking into consideration:  Receiver Based Auto Rate (RBAR)  Opportunistic Auto Rate (OAR) Both use measured SNR of an RTS packet to set the transmission rate for the upcoming data packet in the CTS packet  RBAR targets throughput fairness by allowing one packet transmission per channel access  OAR targets timeshare fairness by allowing a maximum time duration per channel access

30. Multi-Rate Channels

32. Observations Fairness characteristics are somewhat improved as compared to single-rate  Because channel qualities for TA(1) flows are considerably better than for TA(2) flows OAR results in slightly higher throughput than RBAR due to its use of consecutive packet transmissions under high-quality channel conditions

33. Impact of Carrier Sense Range

34. Fairness characteristics are considerably improved, as the impact of hidden terminals and information asymmetry is mitigated  TAP1 is aware of data transmission between TAP3 and TAP4 Disadvantages:  The total goodput is reduced as compared to the case with smaller carrier sense range, since TAP1 and TAP2 are not starved and spatial reuse is inhibited  Not always realistic due to hardware limitations  Reduction of spatial reuse and hence overall throughput

35. Inter-TAP Fairness Algorithm Distributed layer-2 protocol designed to achieve the objectives of the fairness reference model The protocol attempts to eliminate the above starvation and unfairness by limiting flows at the first hop to their system-wide fair rate

36. Inter-TAP Fairness Algorithm Cont. The motivation for a layer 2 solution: It does not require a special-purpose TCP for multi-hop wireless It applies to UDP traffic It can react at faster time scales than end-to-end protocols. The design space for IFA encompasses Classical congestion control issues encountered in wireline networks Issues unique to wireless networks (shared media, hidden terminals, fading channels, etc.) Issues unique to multihop backhaul networks (aggregate fairness granularity, removal of spatial bias, etc.)

37. Baseline UDP/IFA with CSMA and CSMA/CA

38. Observations IFA/CSMA achieves a nearly identical goodput for each TAP of 253 kb/sec to 256 kb/sec, despite the presence of hidden terminals and use of the CSMA MAC IFA does not eliminate the hidden terminal problem Time limiting TAP3 to transmit only 1/6 th of the time  Reduces link layer contention  Provides sufficient spare capacity for the hidden terminal TAP1 CSMA/CA also attains near equal throughput for each TAP at 236 to 238 kb/sec, approximately 7%less than that achieved by CSMA  RTS/CTS overhead due to imperfect media access and collisions

39. Observations Cont. IFA over CSMA and CSMA/CA respectively achieves 76% and 71% of the per-TAP and aggregate throughput as compared to the idealized reference model  The range of the maximum achievable by IEEE 802.11  Achieving higher performance would require reduced collisions or other MAC enhancements Throttling input traffic to its system wide fair time share  High loss due to hidden terminals and contention can be alleviated  Network’s fairness objectives can be approximately achieved

40. TCP/IFA

41. Observations End-to-End congestion control at layer 4 is still required TCP’s end-to-end performance is considerably improved by the IFA algorithm  Prevents starvation of traffic from TAPs 1 and 2 TCP cannot inject bursts of packets in the network, so that the occurrence of excessive losses and timeouts are eliminated TCP does introduce an increased spatial bias as it favors short RTT flows

42. InterTAP Performance Isolation

43. Observations The TCP flow obtains 64% of the idealized objective throughput, whereas the UDP flows obtain 75% IFA ensures that an upstream TCP flow can obtain nearly its fair share

44. Unbalance Number of Flows per TAP

45. Unbalanced Flows Each TAP-aggregated flow should achieve the same time share regardless of the number of mobile users in its collision domain We modify the number of mobile users per TAP  TAP1 and TAP2 each have two MUs transmitting constant-rate UDP traffic  TAP3 has only one MU transmitting TCP traffic The IFA protocol to provide TAP-aggregated fairness as opposed to per flow fairness

46. Scenario with Forward and Reverse Traffic

47. Goodput for Forward and Reverse Traffic

48. Forward and Reverse Traffic Weighted fairness in order to ensure that forward and reverse traffic have the same throughput TAP4 traffic consists of all reverse (or downlink) traffic  It requires a higher weight than TAPs 1 to 3

49. Goodput for Forward and Reverse Traffic For CSMA, downlink traffic is considerably lower as compared to uplink traffic  Hidden terminals The results for CSMA/CA are more balanced  RTS/CTS mitigates effects of hidden terminals The longest path traffic has considerably lower goodput as compared to the shorter path traffic

50. Goodput for Forward and Reverse Traffic IFA achieves better fairness properties The total throughput is lower (9% for CSMA/CA and 13% for CSMA) for this scenario as compared to the scenario with forward traffic only  The available bandwidth decreased because of increased contention due to  The existence of reverse traffic  An increase in the number of transmitting nodes Goodput for Forward and Reverse Traffic

51. Conclusion Multihop wireless backhaul networks have the potential to provide economically viable broadband access networks  Unfortunately, we have shown that current protocols can result in severe unfairness and even starvation of flows farther away from wired Internet entry points IFA achieves better fairness properties Capacity-maximizing strategy can starve multihop flows so we need to make trade off between Fairness and Performance

52. Questions?