1. ZipTx: Harnessing Partial Packets in 802.11 Networks Nate Kushman Kate Ching-Ju Lin, Dina Katabi

2. Wasteful to throw away all correct bytes Current wireless is all-or-nothing A packet that fails the CRC test is discarded

3. Much prior work trys to leverage partial packets [JB07], [WKKS07], [KKBM08], [KKRL03], HARQ, … But….. I only care about my 802.11 laptop

4. How do we build a driver that leverages partial packets? What are the throughput gains in 802.11 networks? – Given 802.11’s auto-rate tries to avoid partial packets Problem

5. Contributions Characterizing the throughput gains of partial packets in 802.11 networks with auto-rate: Indoor WLANS: Auto-rate is effective at eliminating partial packets Gains are limited to 20-25% Challenged outdoor and mobile networks: Auto-rate is ineffective at eliminating partial packets Allows 2-3x throughput gains ZipTx, an 802.11 driver that collects these gains

6. Indoor Testbed 35 nodes Bitrates 6Mb/s  54 Mb/s RSSI 1-34 Outdoor Testbed Measurement Environment 25 nodes Bitrates 1Mb/s  11 Mb/s RSSI 1-12 Method Configure hardware to pass up packets failing the CRC check At each location, cycle between all bitrates For each bitrate, compute correct-packet-throughput and correct-byte-throughput

7. Why Does Rate Adaptation Matter? Say you can’t adapt the bit rate

8. Why Does Rate Adaptation Matter? Say you can’t adapt the bit rate Correct Pkts at 48Mb/s

9. Why Does Rate Adaptation Matter? Say you can’t adapt the bit rate Correct Pkts at 48Mb/s Correct Bytes at 48Mb/s No Rate Adaptation  Dramatic Gains 35x

10. Why Does Rate Adaptation Matter? Correct Pkts at 48Mb/s Correct Bytes at 48Mb/s Say you can adapt the bit rate

11. Gain is about 25%; Much smaller than no autorate Why Does Rate Adaptation Matter? Say you can adapt the bit rate Correct Pkts at 48Mb/s Correct Bytes at 48Mb/s Correct Pkts at 36Mb/s

12. Why Does Rate Adaptation Matter? Say you can adapt the bit rate

13. Why Does Rate Adaptation Matter? Say you can adapt the bit rate

14. Why Does Rate Adaptation Matter? Say you can adapt the bit rate Envelope of Correct Pkts

15. Why Does Rate Adaptation Matter? Say you can adapt the bit rate Envelope of Correct Pkts

16. Why Does Rate Adaptation Matter? Say you can adapt the bit rate Envelope of Correct Pkts

17. Why Does Rate Adaptation Matter? Say you can adapt the bit rate Envelope of Correct Pkts Envelope of Correct Bytes

18. Why Does Rate Adaptation Matter? Say you can adapt the bit rate Rate Adaptation  Limited Gain (about 25%) Envelope of Correct Pkts Envelope of Correct Bytes

19. Where Do the Gains Come From?

26. Partial Packets allow an increase of one bit-rate Where Do the Gains Come From?

27. Is a gain of 25% is all that partial packets can do for 802.11? 25% is the gain in Typical Indoor WLANs But things look different for the more challenged outdoor environment

28. Gains in the Outdoor Environment 2x gain Much larger gains outdoors 3.5x gain

29. Where Do the Outdoor Gains Come From?

30. Most Gains Are Not From Jumping Up Bit-Rates Where Do the Outdoor Gains Come From?

31. In outdoors, auto-rate cannot avoid partial packets

32. What’s the fundamental difference?

33. ZipTx For 802.11 gains vary from 25% up to 3x How do we get them today?

34. Kernel Driver To Hardware ZipTx Modifies the Madwifi Driver Correct Partial Packets Auto-Rate Maximizes Correct bytes From Application

35. Need to recover incorrect bits A software-solution has to choose between: Block-by-block CRC’s Coding How Does ZipTx Correct Partial Packets? But, we don’t know which bits are incorrect! Distributed Errors  Coding is Better

36. How Much Coding? Need to look at percentages of byte-errors in packets 90% 35% correct 60% partial pkts ~5% Erasures Correcting x errors requires 2x parity We need 20% parity Want No Overhead For Fully Correct Packets

37. How Does ZipTx Send Packets? No Ack Still Requires 20% Overhead for 65% of Packets

38. How Does ZipTx Send Packets? 85% have less than 3% byte-errors Send a small amount of parity first, and only if still undecodable send the rest Overhead = 0*.35 +.06*.5 +.2*.15 = 5% Coded packets are piggybacked on next packet

39. Modified Auto-Rate Leverage existing SampleRate algorithm But maximize throughput after correcting partial packets

40. Results

41. Same Indoor/Outdoor testbed Repeatedly choose a source and destination and perform a 1 minute transfer Compared Drivers: – Unmodified Madwifi – ZipTx Metric of comparison: Throughput gain = Experimental Setup

42. Typical Indoor WLANs

43. In indoor WLANs, ZipTx gain is 10-20% Typical Indoor WLANs

44. Outdoor Environment

45. Outdoors, ZipTx average gain is about 2x Outdoor Environment

46. Mobility Experiment Repeatedly, walk down the hall and back

47. Ziptx provides auto-rate algorithm a margin of error Mobility Experiment

48. Related Work Packet Recovery with Soft Information – [JB07], [WSKK07], and [KKBM08] Packet Recover with CRCs – [GJLA06], [DEV05], and [MB05] Packet Recover with Coding – HARQ and [KKRL03]

49. Contributions Auto-rate matters: Typical Indoor: 20-25% Challenged Outdoor: 2x-3x Challenged Mobile: 2x-3x Today’s 802.11 can collect these gains using ZipTx

50. Backup Slides

51. CPU Usage CPU usage is low  practical