1. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 1 VDCF Presentation Greg Chesson, greg@atheros.com Wim Diepstraten, wdiepstraten@agere.com Maarten Hoeben, Maarten.Hoeben@intersil.com Aman Singla, aman@atheros.com Harold Teunissen, hteunissen@lucent.com Menzo Wentink, menzo.wentink@intersil.com

2. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 2 VDCF Overview VDCF is license-free, royalty-free (to be determined) –It is the belief of the proposers this is possible. –There is plentiful prior art. IP statements have been requested. Enhancement to DCF Same state machine as DCF Minimal change to MAC (see document 01/131) Compatible with DCF, PCF Properties Prioritized access to MAC services per Traffic Category (TC) Controls relative bandwidth per TC Controls relative latency and jitter per TC Robust over light, medium, heavy loads Simple Simulation –Extensive validation results (see documents 01/008, 01/133) –Public software: contact authors

3. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 3 VDCF Origins Differentiated Service by traffic category rather than individual flows originates with the IETF Diffserv WG: http://www.ietf.org/html.charters/diffserv-charter.html DQoS proposed by Jan Kruys at San Diego ad hoc meeting in September 2000, captured in later IEEE submissions: Distributed QoS Model for 802.11 (00/267), by Jan Kruys and Harold Teunissen Virtualized DCF access method: Enhance D-QoS through Virtual DCF (00/351), by Maarten Hoeben and Menzo Wentink Baseline D-QoS Proposal (00/399), by Chesson, Diepstraten, Kitchin, Teunissen, Wentink Differentiated Inter-Frame Space, Contention Window, Retry Policy DFWMAC (93/190), by Diepstraten, Ennis, Belanger Priority in CSMA/CA to support distributed Time-Bounded Services (94/058), by Wim Diepstraten Distributed vs Centralized Control Review of Distributed Time Bounded Services (94/121), by Tim Phipps

4. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 4 VDCF Components –Prioritized output queues (queue[i]) –Legacy DCF finite state machine per queue (queue[i]) CWmin differentiated per TC (CWmin[i]), controllable by EAP DIFS differentiated per TC (QIFS[i]), controllable by EAP Queue state machines count backoff slots in parallel Low-priority queues defer to higher-priority queues DCF queue[i] CWmin[i] QIFS[i] Queue[i] TRANSMIT Queue[k] PRI feedback DCF queue[k] CWmin[k] QIFS[k]

5. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 5 Two Controls Contention Window (CW) –Lower-priority TCs select random backoff counters from CWs, on average receiving fewer TxOPS than higher-priority TCs picking from CWs. –Imposes bandwidth and access delay differentiation between TCs –Contention windows expand/contract Local adaptation: binary exponential backoff in response to collision Also controllable by EAP in Beacon CWmin[i] in QoS Parameter Set Element updates aCWmin[i] Inter-Frame Space (IFS) –Different IFS per TC: TxQIFS[i] = SIFS + aQIFS[i] x aSlotTime –Imposes bandwidth and latency differentiation between TCs –Controllable by EAP QIFS[i] in QoS Parameter Set Element updates aQIFS[i]

6. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 6 Why two controls? Both controls provide effective differentiation –CWmin Affects TxOP probability, collision probability average backoff delay –QIFS Low-priority traffic defers to high-priority traffic Slower backoff counting rate for lower-priority traffic Complementary when used together –Use small values for QIFS: e.g. 0, 2, 5, 5 Large QIFS values can exclude traffic –Use smaller range of CWmins; e.g. 15, 15, 31, 63; or 15, 31, 31, 31 –Achieve differentiation with better latency/jitter

7. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 7 Small Examples Load(2,4) => 2 high-priority stations, 4 low-priority stations –Add a station every 3 seconds –Track bandwidth/latency for DCF only CWmin(15,31) and QIFS(0,0) CWmin(15,15) and QIFS(0,1) Load(4,2,10) => 2 high-priority (phone), 4 high-bw (video), 10 background stations –Add a station every 3 seconds: phone, video, background –Then remove a station every 3 seconds –Observe good performance over the entire load range using CWmin(15,15,31) and QIFS(0,2,7)

8. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 8 Bandwidth Differentiation DCFCWmin(15,31) QIFS(0,0)CWmin(15,15) QIFS(0,1) Equal TxOPsSimilar Differentiation

9. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 9 Latency Differentiation DCFCWmin(15,31) QIFS(0,0)CWmin(15,15) QIFS(0,1) High-priority latency plot (per-frame as load increases) 50 ms Hi-pri latency under 20ms Lo-pri latencies Above 50 ms Lower latency Variation with QIFS For guaranteed latency Use HCF

10. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 10 Robust under load changes 4 100 Kbit CBR (phones) 8 Mbit CBR (video) 3 Mbit CBR Background stations Remove Loads

11. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 11 Latency Differentiation Phone latency Video latency 5 ms

12. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 12 Glad you asked that

13. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 13 DCF State Machine IDLETRANSMIT Queue Empty? CCA >= DIFS? Retry Limit? BC=0 Success Fail Abort Retry Ready Yes No Yes No BACKOFF Faithful rendering of Clause 9. Immediate access + post-backoff. See document 01/131 for greater detail. If (CCA>=DIFS) decrement BC

14. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 14 VDCF State Machine (for queue[i]) BACKOFF IDLETRANSMIT Queue[i] Empty? CCA >= QIFS[i]? Retry Limit[i]? BC[i]=0 Success Fail Abort Retry Ready Yes No Yes No PRI OK? Yes No VDCF adds priority test replaces DIFS by QIFS[i], Selects CW from [0,aCWmin[i]]. If (CCA>=QIFS[i] & !Transmit decrement BC[i]

15. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 15 Simulations Simulations based on Berkeley NS2 –Codes simulate full protocol stacks (ARP, UDP, TCP) –Expose protocol stack coupling through AP and other effects See document 01/008 –Demonstrates that priority queues in AP deliver effective QoS in many cases using only legacy DCF –Shows some effects of different retry policies –Shows application of CWmin[i] –Shows effectiveness of PIFS access as it might be used by HCF in the presence of a heavy DCF overload See document 01/133 –Catalog of scenarios with various CWmin[], QIFS[] settings –Incomplete exploration of full parameter space –Demonstrates utility of the controls –Provides starting point for determining default IBSS parameter settings

16. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 16 VDCF Design Choices Distributed Stability Control vs Centralized –Robust: does not depend on EAP or reliable channel for stability –Distributed: self-adapting at station via binary exponential backoff –IBSS-ready: doesn’t need updates for stability Uniform distribution vs Geometric –Better latency variance, delay jitter (see document 01/008) –No “mini-capture effect” (see 01/008) causing backoff amplification Post-backoff/immediate access vs Pre-backoff –Lower latency under light load –Equivalent to Pre-backoff when backlogged queues –Same as legacy DCF Use both QIFS[i] and CWmin[i] –Complementary mechanisms

17. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 17 VDCF Design Choices QoS Parameter Setting vs fast adaptation –QoS Parameters AP adjusts at STA Association time, or RSVP time AP adjusts to observed load average – not time-critical “slow” adaptation: unlikely to stimulate control oscillation –Fast Adaptation Unacknowledged broadcast not well-suited for wireless media System adaptation rate (sample+decide+broadcast+adopt) slower than rate of change of offered load in many cases: cause of oscillation, degradation. Fast adaptation consumes bandwidth, TxOPs, MAC logic cycles Independent queue[i] state –Fairness across TCs and stations –Backoff counts (BC[i]) retain age ordering (i.e. ensure forward progress)

18. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 18 Implementation Factors Retain power-of-2 CWmin intervals –Simple arithmetic, no division/mod ops needed –Simple random number generation Random number generation rate –Once per TxOP per queue Must recognize QoS-DATA and (TBD) TCID tags –Otherwise no new frame exchange sequences One new information element to process: QoS Parameter Set –Appears in Beacon and Probe Response –Adjusts CWmin[] and QIFS[] values Sequence numbers –No change at sender, can assign sequence number at TxOP –Receive cache must include TC, i.e. triples instead of tuples.

19. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 19 Implementation Factors New MIB variables: aCWmin[0-7], aQIFS[0-7], aSSRC[0-7], aSLRC[0-7], aCWmax[0-7]. State variables for each output queue: Backoff Counter BC[i] Short/long retry counters QSRC[i] and QLRC[i] Contention window CW[i] Virtual collisions Priority test applied when BC[i] reaches zero Losing queue[i] goes into backoff state Backoff Frame ordering is not preserved between TCs All queues can be in backoff at the same time A single counter (plus logic) can represent multiple backoff states

20. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 20 Summary Simple –minimal control mechanism Safe –builds on proven MAC Differentiated Service –bandwidth differentiation –latency differentiation and mitigation Robust –self-adaptive, but also controllable –differentiates over changing loads

21. doc.: IEEE 802.11-01/132r1 Submission March 2001 Greg Chesson et al, Atheros Slide 21 Conclusion Simple is good