1. DOMINO: Relative Scheduling in Enterprise Wireless LANs Wenjie Zhou (Co-Primary Author), Dong Li (Co-Primary Author), Kannan Srinivasan, Prasun Sinha 1

2. Enterprise Networks 2 Important research topics: -Channel assignment -AP association -Power adaptation -Channel access Internet Router AP 1 AP 2 AP N … Client 1 Client 2 … Client 3 Client N

3. 3 Channel Access Schemes in enterprise networks: - Distributed Coordination Function (DCF) - Downlink-only Centralized Schemes - Fully Centralized Schemes

4. Distributed Channel Access: DCF (WiFi) 4 Pros: - Simple to implement - Robust to failures AP 2 AP 1 AP 3 C2C2 C1C1 C3C3 Exposed Hidden Cons: - Hidden and exposed terminal problems - Low efficiency Interfering nodes Flow direction

5. Downlink-only Centralized Schemes CENTAUR (Mobicom’09): Downlink packets that could be sent simultaneously are forwarded to the APs at the same time. OmniVoice (MobiHoc’11): Downlink packets are sent according to broadcast schedules. 5 Pros: - No modification on clients Cons: - Downlink traffic only

6. Fully Centralized Scheme Schedule both uplink and downlink traffic 6 AP 1 -> C 1 C 2 -> AP 2 AP 3 -> C 3 Overall ~61% AP 2 AP 1 AP 3 C2C2 C1C1 C3C3 Interfering nodes Flow direction

7. Domino a practical platform to enable arbitrary centralized scheduling algorithms without requiring tight time-synchronization 7

8. DOMINO Outline Rapid OFDM Polling (ROP) – Obtain the queue status of clients for uplink scheduling Relative scheduling – Avoid tight time synchronization Schedule converter – Create schedules for relative scheduling 8

9. Central controller AP 1 AP 2 AP N … Client 1 Client 2 … Client 3 Client N Design Overview 9 Collector (ROP) Scheduler Converter queue size time schedule relative schedule

10. DOMINO Outline Rapid OFDM Polling (ROP) – Obtain the queue status of clients for uplink scheduling Relative scheduling – Avoid tight time synchronization Schedule converter – Create schedules for relative scheduling 10

11. Question: How can we collect the queue status of clients efficiently? Solution: Concurrent transmission based on Orthogonal frequency- division multiplexing (OFDM) Central controller AP 1 AP 2 AP N … Client 1 Client 2 … Client 3 Client N ROP: Rapid OFDM Polling 11

12. ROP: Rapid OFDM Polling Clients transmit queue status using subcarriers 12 Client 1Client 2 Practical issues: -Freq offset -Time offset -Power mismatch (details in paper) Related work: -PAMAC (INFOCOM’09) -B2F (MobiCom’11)

13. Subcarrier Separation 13 111111 0 separation 3 sub separation Detection threshold TX 1 RX TX 2 Experiment results suggest 3 sub separation is enough

14. 14 ROP collects the queue status of all clients with little overhead: - 40 μs (polling message) + 16 μs (OFDM symbol) - regular packet duration: 1000 μs - multiple regular transmissions/poll

15. DOMINO Outline Rapid OFDM Polling (ROP) – Obtain the queue status of clients Relative scheduling – Avoid tight time synchronization Schedule converter – Create schedule for relative scheduling 15

16. 16 AP 1 ---> C 1 : AP 4 ---> C 4 : AP 2 ---> C 2 : AP 3 ---> C 3 : Data PacketACK Data PacketACK Data Packet ACK Misalignment Collision ! AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4 Slot 1 AP 1 ---->C 1 AP 4 ---->C 4 μs level synchronization required One Wi-Fi slot: 9 μs Why time synchronization? Interfering nodes AP-client association Currently transmitting Slot 2 AP 2 ---->C 2 AP 3 ---->C 3

17. Current Time Synchronization Scheme Network Time Protocol (NTP), Precision Time Protocol (PTP), Reference-Broadcast Synchronization (RBS) (SIGOPS’02), Sourcesync (SIGCOMM’10): – Low accuracy; Or – Expensive hardware; Or – Low accuracy in large network. 17

18. 18 Can we avoid tight time synchronization?

19. Relative Scheduling 19 AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4 Slot 1Slot 2 AP 1 ---->C 1 AP 4 ---->C 4 AP 2 ---->C 2 AP 3 ---->C 3 Data PacketACK Data PacketACK Data PacketACK Data PacketACK Interfering nodes AP-client association Currently transmitting AP 1 ---> C 1 : AP 4 ---> C 4 : AP 2 ---> C 2 : AP 3 ---> C 3 :

20. Relative Scheduling 20 AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4 Slot 2Slot 3 AP 1 ---->C 1 AP 4 ---->C 4 AP 2 ---->C 2 AP 3 ---->C 3 Data PacketACK Data PacketACK Data PacketACK Data PacketACK Transmission alignment achieved Interfering nodes AP-client association Currently transmitting AP 1 ---> C 1 : AP 4 ---> C 4 : AP 2 ---> C 2 : AP 3 ---> C 3 :

21. Node signatures as triggers: – A sequence of bits with a certain length – These sequences are orthogonal to each other – High detecting ratio even under interference – Experiment results: 4 combined signatures can be decoded correctly 4 transmissions can be triggered by one node 21

22. 22 Only APs know the schedules from the central controller How can we ask the clients to send the triggers?

23. 23 Data PacketS A3 S A2 A1:A1: C1:C1: ACK S A3 SIFS 1 slot S′ The combined signatures that should be sent by the client The combined signatures that should be sent by the AP A special signature that notifies the start of transmission A2A2 A1A1 A3A3 C2C2 C1C1 C3C3

24. A2A2 A1A1 A3A3 C2C2 C1C1 C3C3 24 Data Packet S A2 A1:A1: C1:C1: S A3 S A3 SIFS 1 slot S′ ACK

25. DOMINO Outline Rapid OFDM Polling (ROP) – Obtain the queue status of clients Relative scheduling – Avoid tight time synchronization Schedule converter – Create schedule for relative scheduling 25

26. Schedule Converter Requirements: – Every transmitter should be triggered – Polling packets should also be scheduled – Backup triggers should be included in case of transmission failure – Details in paper 26 Arbitrary ScheduleRelative Schedule ?

27. Experiment 27 >3X >1.5X USRP

28. Trace Driven Simulation Simulation Setup: – RSS trace collected from a 40 Wi-Fi nodes testbed – Randomly picked 10 APs and 2 clients per AP Other schemes: −CENTAUR: Downlink traffic scheduled; using fixed backoff to align transmission −DCF: 802.11 standard (Wi-Fi) 28

29. UDP Throughput & Delay 29 Downlink traffic only 1.74X 0.5X

30. UDP Throughput & Delay 30 Uplink and downlink traffic 1.24X Heavy tail Low fairness

31. TCP Throughput 31 Downlink traffic only 1.15X TCP ACK as regular packet

32. Conclusions Domino: a platform to enable centralized scheduling algorithms without requiring tight time-synchronization: – Queue information of clients are efficiently collected using one OFDM symbol – Nodes transmit relatively one after another instead of according to time stamps Future work: coexistence with existing Wi-Fi 32

33. Backup slides 33

34. 34 Why is CENTAUR behaving worse than DCF?

35. UDP Throughput & fairness 35 - 24%-74% throughput gain - High and stable fairness

36. TCP Throughput & fairness 36 - 10%-15% throughput gain - TCP ACK as regular packet - High and stable fairness

37. Evaluation: UDP & TCP Delay 37 DCF: 2X higher - Queuing delay Similar delay performance: - Queuing delay - TCP congestion control

38. UDP Throughput & Delay 38 Uplink and downlink traffic 1.24X Heavy tail 2X

39. TCP Throughput & Delay 39 Uplink traffic only 1.15X

40. Domino Solution Overview 40 Contention Free PeriodContention Period Slot 1Slot 2Slot 3 … Slot N AP 1 → C 1 AP 2 → C 2 … AP M → C M Concurrent transmissions

41. Evaluation: Misalignment 41 Alignment achieved - Slot size: 9 μs > 2 μs

42. Current Time Synchronization Scheme Network Time Protocol (NTP): time accuracy of about 1ms in a quiet Ethernet network. Precision Time Protocol (PTP): requires specialized and expensive hardware. Reference-Broadcast Synchronization (RBS) (SIGOPS’02): synchronization accuracy decreases with network size. Sourcesync (SIGCOMM’10): one collision domain 42

43. Throughput Gain of network with 80 nodes 43 ~58%

44. ROP: Rapid OFDM Polling Client TX queue state over subcarriers Polling strategy 44 0 1 39 100 109 127-128 -109- 100 -9-3 2 -2 4 -4 DCDC ………………… … guard band subchannel 0 subchannel 11subchannel 12 subchannel 23 guard subcarriers Polling Packet From AP Client 0 Client 1 Client 2 Client N … Client 3 1 slot Subchannel Time 0 1 2 3 … N FFT window CP

45. ROP: Rapid OFDM Polling 45 Polling Packet From AP Client 0 Client 1 Client 2 Client N … Client 3 1 slot Subchannel Time 0 1 2 3 … N FFT window CP

46. DOMINO: Example 46 AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4

47. Subcarrier Separation 47 (38dB, 99.9%, 3 sub) Tradeoff: - Less overhead - Higher decoding ratio TX 1 RX TX 2

48. Relative Scheduling: Node signatures as trigger Node signature are orthogonal to each other, easier to detect. 48

49. Existing work: MIFI 49

50. Existing work: CENTAUR 50

51. Domino Solution Overview 51

52. ROP: How it performs 52 Decoded OFDM symbols of two clients at adjacent subchannels with guard interval. (30db diff. RSSI at AP)

53. Relative Scheduling 53 AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4 Slot 1Slot 4Slot 3Slot 2 AP1---->C1 AP4---->C4 AP2---->C2 AP3---->C3 AP1---->C1 AP4---->C4 AP2---->C2 AP3---->C3 AP 1  C 1 AP 2  C 2 AP 3  C 3 AP 4  C 4

54. Relative Scheduling: Node signature as trigger Node signature are orthogonal to each other, easier to detect. – AP to client – Client to AP 54 Data PacketS1S1 S2S2 AP: Client: ACK S1S1 SIFS 1 slot S′ Data Packet ACK S2S2 AP: Client: S1S1 S1S1 SIFS 1 slot S′

55. Relative Scheduling: How it performs 55

56. Schedule Converter Inbound & outbound Constraint – 1<= Inbound <=2 – Outbound <=4 56 In bound Sender Receiver Out bound Arbitrary ScheduleRelative Schedule ?

57. Schedule Converter Insert Fake Link – Saturate the network with fake links at each slot Retain Last Slot – Last slot of current schedule is used as the first slot of next schedule Insert Polling slot – Insert polling slots between slots 57

58. Evaluation: Experiment 58 USRP

59. Evaluation: Simulation 59

60. Evaluation: Misalignment 60

61. Evaluation: UPD & TCP 61

62. Discussion Triggering may not easy Is fixed packet size good? Building conflict graph dynamically Low traffic, low efficiency 62 AP 1 C1 AP 2 C2

63. Challenge : Time Synchronization Network Time Protocol (NTP): time accuracy of about 1ms in a quiet Ethernet network. Precision Time Protocol (PTP): require specialized and expensive hardware. Reference-Broadcast Synchronization (RBS): accuracy decreases as network size increases. 63

64. Centralized Scheme Schedule both uplink and downlink traffic 64 AP 2 AP 1 AP 3 C2C2 C1C1 C3C3

65. DOMINO: Example 65 9291939490 Batch 10 Batch 11 Slots Links AP 1 C1C1 C1C1 AP 2 C2C2 C2C2 AP 3 C3C3 C3C3 AP 4 C4C4 C4C4 0 123 Batch 0 2 fake 3 AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4 Polling packet Data packet 1

66. Distributed Channel Access: DCF 66 Pros: - Simple to implement - Robust to failures AP 2 AP 1 AP 3 C2C2 C1C1 C3C3 Exposed Hidden Cons: - Hidden and exposed terminal problems - Low efficiency

67. Time Synchronization 67 AP 1 C1C1 AP 2 C2C2 AP 3 C3C3 AP 4 C4C4 Slot 1Slot 4Slot 3Slot 2 AP 1 ---->C 1 AP 4 ---->C 4 AP 2 ---->C 2 AP 3 ---->C 3 AP 1 ---->C 1 AP 4 ---->C 4 AP 2 ---->C 2 AP 3 ---->C 3

68. Enterprise Network 68 Internet Central Server AP 1 AP 2 AP N … Client 1 Client 2 Client 3 Client M … Control Plane: - Channel Assignment - Client Association - Power control What about data plane?

69. Exposed and hidden terminals 69 C2C2 AP 2 C1C1 AP 1 C3C3 AP 3 Exposed Hidden Centralized schedule could avoid hidden terminals while utilize exposed terminals

70. Expected Improvement DCF : purely distributed CENTAUR : half-distributed, half centralized DOMINO: centralized 70 centralization

71. Design Overview 71 Internet Central Server AP 1 AP 2 AP N … Client 1 Client 2 Client 3 Client M … Collector Scheduler Converter queue size time schedule relative schedule Obtaining clients queue status Identifying exposed and hidden links Time synchronization Challenges: