Fast Resilient Jumbo Frames in Wireless LANs

Transcript

1. Fast Resilient Jumbo Frames in Wireless LANs Apurv Bhartia University of Texas at Austin apurvb@cs.utexas.edu Joint work with Anand Padmanabha Iyer, Gaurav Deshpande, Eric Rozner and Lili Qiu IWQoS 2009 July 15, 2009
2. 2 Motivation •  Lossy wireless medium •  Novel techniques have been proposed … … but each of them alone is insufficient Partial Recovery Jumbo Frames Rate Adaptation Our goal: identify the synergy between these techniques and exploit it
3. 3 State of the Art •  Jumbo Frames –  Proprietary solutions for frame aggregations [Atheros Super G, TI frame concatenation] –  802.11n frame aggregation standard •  Require specific hardware support •  Entire packet needs to be retransmitted •  Partial Packet Recovery –  Require specific hardware support [MRD, SOFT, PPR] –  Leverage PHY layer information [SOFT, PPR] •  if PHY layer information is available, FRJ can benefit to provide higher gain •  Rate Adaptation –  SampleRate, ONOE (madwifi), RRAA –  Over-estimates the actual loss rate •  Adapt rate according to frame loss rate •  Over-estimates the actual loss rate Holistic Approach is missing !
4. 4 Our Contributions •  Identify interactions between the three techniques –  Exploit the synergy between the schemes –  Works for both single and multi-hop topologies •  Develop resilient jumbo frames –  Achieve high throughput under both low and high loss conditions •  Develop partial recovery aware rate adaptation •  Develop a prototype implementation
5. 5 Synergy Between Design Space Partial Recovery Jumbo Frames Rate Adaptation Constant MAC overhead Reduces relative cost of RTS/CTS Loss Increases with frame size Increases effectiveness of jumbo frames Less collisions – effective recovery Higher tx rates! Increased tx rates reduces contention losses Reduces effective data loss rate Better partial recovery Higher tx rates – increases relative MAC overhead More data for constant overhead Benefit increases with increased tx rates Partial Recovery Aware Rate Adaptation Partial Recovery Aware Rate Adaptation Partial Recovery Aware Rate Adaptation
6. 6 Resilient Jumbo Frames S R •  Use jumbo frames –  High throughput in good conditions –  In bad conditions … •  … re-transmit only corrupted segments –  Saves the overhead of retransmitting complete frames 2.5 ACK
7. 7 Resilient Jumbo Frame •  Data Frames •  Core Components –  Resilient Jumbo Frames which applies partial recovery to jumbo frames –  Partial recovery ‘aware’ rate adaptation Header 4 4 4 4 4 4 1 1 2 2 Segment 1 CRC Segment 2 CRC Segment N CRC Frame ID Type Rate Bitmap SS Header CRC Length
8. 8 Resilient Jumbo Frame (Cont.) •  Receiver Feedback –  Combination of MAC-layer and 2.5-layer ACKs –  MAC-layer ACKs •  Adjustment of back-off window in IEEE 802.11 •  Increased reliability and efficiency than 2.5 ACKs –  2.5-layer ACKs •  To support partial recovery •  Unicast for improved reliability and cumulative Frame Offset Segment Bitmap 1 Frame CRC Header Frame Offset N Segment Bitmap N Start Frame Seg No Type Rate Frame Bitmap
9. 9 Approach •  Retransmission –  Disable MAC layer retransmissions •  set MAC retry count = 0 •  Retransmit the frames at the 2.5-layer –  Triggered by •  2.5-layer ACKs –  If 1st Retx: frames with higher seq nos or some segments in this frame are ACKed [first data transmissions is in-order] –  If 2nd or higher: some new segments in this frame are ACKed •  Retransmission Timeout –  Standard approach as in TCP
10. 10 Partial Recovery Aware Rate Adaptation –  Traditional schemes identify optimal rate using frame loss rate •  Overestimates the loss rate •  Lower data transmissions rates are selected –  Challenges for the ‘new’ scheme •  Accurate estimation of channel condition at various data rates •  Selecting rate that maximizes throughput under partial recovery Estimate throughput based on loss statistics !
11. 11 Partial Recovery Aware Rate Adaptation •  Estimating Channel Condition –  Sender periodically broadcasts probe packets –  Sent at different data rates •  CurrRate r [current data rate] •  CurrRate- r [one rate below the current data rate] •  CurrRate+ r [one rate above the current data rate] –  Sent at a frequency of 5 probes/second •  Limit the overhead Type Payload Probe ID Rate Header CRC Per rate
12. 12 Partial Recovery Aware Rate Adaptation •  Probe Response –  Sent by the receiver –  Estimates the channel condition using •  Header Loss Rate (HL) – header corruption •  Segment Loss Rate (SL) – segment corruption •  Communicates this info using probe response –  Transmitted via MAC-layer unicast •  High reliability –  Default Probe response [HL = 1, SL = 1] •  To account for lost probes Type Probe Response ID Rate1 Frame CRC BER1 HL1 Rate1 BER1 HL1
13. 13 Partial Recovery Aware Rate Adaptation •  Sender selects the rate that gives the best throughput estimation T = ∑ Pi × (Backoff + DIFS + i=1..MaxRetries + 1 DATA + SIFS + ACK + useRTS + RTSOverhead ) preambleTime + (HS + NSi + segmentSize) rate Pi = 1 i = 1 Pi-1 × (HL + (1 – HL) × (1- (1 – SL) )) otherwise NSi-1 Throughput = (NS1 – NSMaxRetries + 2 ) × SegmentSize/T NSi = 30 i = 1 NSi-1 × (HL + (1 – HL) × SL ) otherwise RTS + SIFS + CTS + SIFS NSi Probability of sending the ith tx Time for ith data tx No of segments in ith tx
14. 14 Testbed Topology •  24 machines •  Madwifi driver and CLICK toolkit •  Initial rate = 24Mbps •  Tx Power = 18 dBm Total throughput Per flow throughput Jain’s Fairness Index
15. 15 Schemes Compared •  Sample Rate using 1500 byte frames [SR/ 1500-bytes] •  Sample Rate using 3000 byte frames [SR/ 3000-bytes] –  Same as SR/1500, but uses jumbo frames –  Similar to Atheros Super G Fast Frame feature •  FRJ using 3000 byte frames, 30 segments With and without RTS/CTS
16. 16 Experimental Results: Single Flow Throughput (Mbps) Cumulative Fraction SR/1500: 0.68 Mbps SR/3000: 0.68 Mbps FRJ: 1.1 Mbps SR/1500: 14.17 Mbps SR/3000: 16.93 Mbps FRJ: 23.81 Mbps Moderate Link Conditions: Partial Recovery is more effective FRJ benefit is 40.6% - 68.0% under single flow
17. 17 Experimental Results: Multiple Flows 0 5 10 15 20 25 -5 1 2 4 6 8 # Flows Average Total Throughput (Mbps) FRJ SR/ 1500 bytes SR/3000 bytes FRJ w/ RTS SR/1500 bytes w/ RTS SR/3000 bytes w/ RTS Schemes w/o RTS/CTS perform well Randomly chosen flows! FRJ constantly outperforms More collisions => increase in header losses FRJ benefit ranges from 10% (1 flow) to 64% (6 flows)
18. 18 Experimental Results : Multiple Flows Throughput (Mbps) Cumulative Fraction Average Throughput SR/1500: 0.84 Mbps FRJ: 1.68Mbps SR/3000: 1.05 Mbps SR/1500: 0.30 Mbps SR/3000: 0.38 Mbps FRJ: 0.57 Mbps
19. 19 Experimental Results: Multiple Flows •  Fairness –  Difference is within 10% –  Most cases it is close to 0 # Flows Fairness Index FRJ’s performance gain does not come at the cost of compromising fairness!
20. 20 Conclusion •  Main contributions –  Identify interplay between jumbo frames, PPR and rate adaptation •  Jumbo frames with partial recovery •  Partial recovery aware rate adaptation –  Demonstrate the effectiveness of this solution through testbed experiments •  Future work –  More effective partial recovery schemes and coding techniques –  Dynamically configurable RTS/CTS –  FRJ-aware route selection
21. Thank you! apurvb@cs.utexas.edu
22. 22 0 5 10 15 20 25 -5 1 2 4 6 8 # Flows Average Total Throughput (Mbps) FRJ SR/ 1500 bytes SR/3000 bytes FRJ w/ RTS SR/1500 bytes w/ RTS SR/3000 bytes w/ RTS