1. Packet Mixing: Superposition Coding and Network Coding Richard Alimi CS434 Lecture Joint work with: L. Erran Li, Ramachandran Ramjee, Harish Viswanathan, Y. Richard Yang 2/12/2009

2. 2 Wireless Mesh Networks City- and community-wide mesh networks widely used  New approach to the “last mile” of Internet service  In United States alone [muniwireless.com, Jan 16, 2007] : 188 deployed 148 in-progress or planned

3. 3 Mesh Network Structure APs deployed, some connect directly to Internet  Street lamps, traffic lights, public buildings Clients associate with nearest AP Traffic routed to/from Internet via APs (possibly multi-hop)

4. 4 Limited Capacity of Mesh Networks Current mesh networks have limited capacity [Li et al. 2001, dailywireless.org 2004] Increased usage will only worsen congestion  More devices  Larger downloads, P2P, video streaming  Limited spectrum Network-wide transport capacity does not scale [Gupta and Kumar 2001] Must bypass traditional constraints

5. 5 Packet Mixing for Increased Capacity Multiple packets transmitted simultaneously  Same timeslot  Cross-layer coding techniques  No spreading (unlike CDMA) Receiver(s) decode own packets  Possibly use side-information (e.g., packets previously overheard)

6. 6 Packet Mixing Objective Objective  Construct a mixed packet with Maximum effective throughput Sufficiently-high decoding probability at receivers Mixture of coding techniques Currently consider two techniques  Downlink superposition coding  XOR-style network coding

7. 7 Recap: Physical Layer Signal Modulation Signal has two components: I and Q Represented on complex plane Sender  Map bits to symbol (constellation point) Receiver  Determine closest symbol and emit bits 00 01 11 10 QPSK

8. 8 Downlink Superposition Coding Basic idea  Different message queued for each receiver  Transmit messages simultaneously  Exploit client channel diversity Example

9. 9 Downlink Superposition Coding Basic idea  Different message queued for each receiver  Transmit messages simultaneously  Exploit client channel diversity Example “01” “11” Stronger receiver Layer 2 “High resolution” Weaker receiver Layer 1 “Low resolution”

10. 10 Downlink Superposition Coding Basic idea  Different message queued for each receiver  Transmit messages simultaneously  Exploit client channel diversity Example “01” “11” Stronger receiver Layer 2 “High resolution” Weaker receiver Layer 1 “Low resolution”

11. 11 SC Decoding: Successive Interference Cancellation “01” “11” Weaker Receiver Decode Layer 1

12. 12 Stronger Receiver SC Decoding: Successive Interference Cancellation (1) Decode Layer 1 “01” “11”

13. 13 Stronger Receiver SC Decoding: Successive Interference Cancellation (2) Subtract and decode Layer 2 “01” “11” (1) Decode Layer 1

14. 14 SC Quantization Gains Discrete rates are common in standards (including 802.11) In other words...  Channel qualities are quantized distance 36 Mbps 24 Mbps

15. 15 SC Quantization Gains Idea  Steal extra power from one receiver without affecting data rate  Allocate extra power to second layer destined for a stronger receiver Use largest rate achievable with this extra power Can derive formula for computing these gains: where:

16. 16 SC Quantization Gains in 802.11a/g Distances  R 1 : varies  R 2 : 40 meters Path-loss Exponent: 4 Noise: -90 dBm Remaining parameters consistent with Cisco Aironet 802.11g card

17. 17 SC Scalability Analysis Recall capacity analysis for arbitrary network:  Radio Interface constraint  Interference constraint First we'll show an upper bound:

18. 18 SC Scalability Analysis Without Superposition Coding bit time t 1 2 3 4 5 1 2 3 4 5 With Superposition Coding

19. 19 SC Scalability Analysis Without Superposition CodingWith Superposition Coding bit time t We now have up to n-1 transmissions per bit-time! 1 2 3 4 5 1 2 3 4 5

20. 20 SC Scalability Analysis Without Superposition CodingWith Superposition Coding Up to n-1 concurrent transmissions, so...

21. 21 SC Scalability Analysis Without Superposition Coding Up to n-1 concurrent transmissions, so... Total area due to interfering transmissions reduces by factor of n With Superposition Coding

22. 22 SC Scalability Analysis Changes to capacity analysis  At most n-1 concurrent transmissions (superposition coding with n-1 layers)  # of interfering transmissions reduced by a factor of n Modified constraints produce O(n) upper bound Is the upper bound achievable? Yes

23. 23 XOR-style Network Coding Basic idea  Nodes remember overheard and sent messages  Transmit bitwise XOR of packets:  Receivers decode if they already know n-1 packets Example 1 2n 21 21 R1R1 R2R2

24. 24 Putting it Together: Packet Mixing 4 Flows  Packet d has dest R d Without packet mixing  8 transmissions required With packet mixing  5 transmissions required Overhearing linkRouting link R4R4 R1R1 R3R3 R2R2 1 2 3 4

25. 25 Putting it Together: Packet Mixing R4R4 R1R1 R3R3 R2R2 1 2 3 4 1 1 (1) R 2 sends Pkt 1 to AP Overhearing linkRouting link

26. 26 Putting it Together: Packet Mixing R4R4 R1R1 R3R3 R2R2 1 2 3 4 1 1 2 2 (1) R 2 sends Pkt 1 to AP (2) R 4 sends Pkt 2 to AP Overhearing linkRouting link

27. 27 Putting it Together: Packet Mixing R4R4 R1R1 R3R3 R2R2 1 2 3 4 1 1 2 2 3 3 (1) R 2 sends Pkt 1 to AP (2) R 4 sends Pkt 2 to AP (3) R 1 sends Pkt 3 to AP Overhearing linkRouting link

28. 28 Putting it Together: Packet Mixing R4R4 R1R1 R3R3 R2R2 1 2 34 1 1 2 2 3 3 4 4 (1) R 2 sends Pkt 1 to AP (2) R 4 sends Pkt 2 to AP (3) R 1 sends Pkt 3 to AP (4) R 3 sends Pkt 4 to AP Overhearing linkRouting link

29. 29 Putting it Together: Packet Mixing R4R4 R1R1 R3R3 R2R2 1 2 34 1 1 2 2 3 3 4 4 (1) R 2 sends Pkt 1 to AP (2) R 4 sends Pkt 2 to AP (3) R 1 sends Pkt 3 to AP (4) R 3 sends Pkt 4 to AP (5) AP sends mixed packet using SC: (a) Layer 1: (b) Layer 2: 12 34 Overhearing linkRouting link

30. 30 Scheduling under Packet Mixing Per-neighbor FIFO packet queues  Q d is queue for neighbor d – first denoted by head(Q d ) Total order on packets in all queues  Ordered by arrival time – first denoted by head(Q) Rule: always transmit head(Q)  Prevents starvation d2d2 d3d3 d1d1 head(Q) head(Q 3 )

31. 31 Superposition Coding Scheduling Overview  Layer 1: Select head(Q) with dest d 1 at rate r 1  Layer 2: Select floor(r 2 / r 1 ) packets for dest d 2 ≠ d 1 at rate r 2  Allows different rates for each layer Selects best rates given current channels Ensures sufficient decoding probabilities Destination 2, rate 12 Mbps head(Q 2 ) 3 packets from Q 4 Layer 1 Layer 2 Destination 4, rate 36 Mbps Rate:12 Mbps Packets:4 Throughput:48 Mbps

32. 32 Superposition and Network Coding Scheduling Algorithm  Iterate over discrete rates for each layer, r 1 and r 2 Layer 1: Select network-coded packet, N 1 packets encoded Layer 2: Selects floor(r 2 / r 1 ) network-coded packets, N 2 packets encoded  Only consider neighbors that support r 2 in second layer Effective throughput is r 1 (N 1 + N 2 ) head(Q 1 ) head(Q 2 ) head(Q 3 ) head(Q 4 ), head(Q 5 ) head(Q 6 ), head(Q 7 ) head(Q 8 ) Layer 1 Layer 2

33. 33 Evaluations: Setup Algorithms implemented in ns-2 version 2.31 Careful attention to physical layer model  Standard ns-2 physical layer model does not suffice  Use packet error rate curves from actual 802.11a measurements [Doo et al. 2004]  Packet error rates used for physical layer decoding and rate calculations Realistic simulation parameters  Parameters produce similar transmission ranges as Cisco Aironet 802.11g card in outdoor environment

34. 34 Evaluations: Network Demand Setup  1 AP  10 clients  8 flows  Vary client sending rate Packet mixing gains are sensitive to network demand Queues are usually empty with low demand  Few mixing opportunites Network Coding shows ~3% gain with TCP [Katti et al. 2006]

35. 35 Evaluations: Internet → Client Flows Setup  1 AP  20 clients  16 flows  Backlogged flows  Vary % of flows originating at AP Superposition Coding superior when Internet → client flows are common Throughput gains as high as 4.24

36. 36 GNU Radio Implementation Open Source software radio  RF frontend hardware (USRP)  Signal processing in software Components  Implementation of Superposition Coding in GNU Radio environment  802.11 MAC implemented with Network Coding support Measurement Results

37. 37 Backup Slides

38. 38 Example Coding Techniques Transmitter-side  Downlink superposition coding [Cover 1972, Bergmans and Cover 1974]  XOR-style network coding [Katti et al. 2006] Receiver-side  Uplink superposition coding  Analog [Katti et al. 2007] and physical-layer [Zhang et al. 2006] network coding

39. 39 Multirate NC: mnetcode SC requires multirate for better gains, so extend NC as well Algorithm  Run single-rate COPE algorithm snetcode(r) for each rate r and select best Skip rate if not supported by neighbor Only consider head(Q d ) for each neighbor d N packets XOR'd at rate r Effective throughput is N r

40. 40 Simple Cross-layer Mixing Utilize physical- and network-layer coding Algorithm  SC Layer 1: Select NC packet with mnetcode Must include head(Q)  SC Layer 2: Select packet with G opp Problems  No NC used in Layer 2 packets  Limited rate combinations head(Q 1 ) head(Q 2 ) 3 packets from Q 4 Layer 1 Layer 2

41. 41 Evaluations: Client → Client Flows Setup  1 AP  20 clients  Backlogged flows  Vary # of flows Both SC and NC mixing alone improve with # of flows  More opportunities Gains each SC and NC exploited successfully by SC1 and SCJ schedulers