Presentation for Proposed p-DCF Contention Access Enhancement

1. Presentation for Proposed p-DCF Contention Access Enhancement Jin-Meng Ho, Sid Schrum, Khaled Turki Donald P. Shaver and Matthew B. Shoemake Texas Instruments Incorporated 12500 TI Blvd. Dallas, Texas 75243 (214) 480-1994 (Ho) jinmengho@ti.com

2. Outline • P-DCF uses one backoff timer per station, just like legacy DCF. • It does not stack multiple DCFs within each station. • It does not have the issue of checking, and resolving, simultaneous expiration of multiple backoff timers at any given station. • P-DCF separates external behavior (access to medium) from internal behavior (selection from queues). • Each ESTA performs its external contention just as a legacy DCF station. • Multi-priority service per station appears as an internal enhancement to the legacy DCF MAC. • P-DCF obeys DIFS usage as specified for legacy DCF. • No extra tiers of contention are required. • No new backoff countdown rules are specified.

3. Performance • P-DCF achieves differentiated service for prioritized traffic. • Higher-priority data encounters smaller access delay than lower-priority data. • Lower-priority traffic is not starved nor suffers excessive access delay. • P-DCF improves access delay over legacy/stacked DCF. • Higher-priority frames under P-DCF experience less delay than best-effort frames under legacy DCF. • Not a case with stacked DCF. • Lower-priority frames under P-DCF experience the same delay as best-effort frames under legacy DCF. • Lower-priority frames are starved under stacked DCF. • P-DCF increases channel throughput compared to legacy/stacked DCF. • More bits per second can be transmitted per channel. • More stations and streams can be served per BSS.

4. Coordinated Contention • Traffic Category Permission Probabilities (TCPPs) • Each traffic category (TC) is assigned a TCPP. • A frame from TCi is transmitted with a probability = TCPPi (conceptually). • A frame from a WSTA is sent with a permission probability (PP) equal to the sum of the latest TCPPs for active local TCs (conceptually). • Infrastructure Network • The hybrid coordinator (HC) regularly updates TCPPs for TCs of eight priorities and broadcasts them via an ECA Parameter Set element in beacons. • Update intervals of 20 - 50 ms long are found to provide good performance. • IBSS and Backup Contention Access • TCPPs are adjusted in a way similar to binary exponential backoff for DCF. Element ID (12) Length (8) TCPP Values (TCPP0, …, TCPP7) (8 octets) ECA (Enhanced Contention Access) Parameter Set

5. Distributed Contention • IBSS and Backup Contention Access • If an active WSTA is located in an IBSS, or if it has not received TCPP values for 50 TUs from the HC, it performs its contention for a frame transmission using the TCPP values calculated on its own: • Any active local TC that has a non-zero TCPP value continues to have the same TCPP value until it has a frame transmitted. • A local TC of priority i that has just successfully sent a frame has a TCPP value equal to TCPPi, max, where TCPP0, max = 1/33, and TCPPi, max = 2/17, i = 1, 2, …7, if the channel is busy, and has a TCPP value equal to TCPPi,idle, where TCPP0, idle = 1/4, and TCPPi, idle = 1/2, i = 1, 2, …7, if the channel is idle. • A local TC of priority i that has a retried frame to send after a collision changes its TCPP value from TCPPi to max [TCPPi,min, 2  TCPPi / (4 – TCPPi)], where TCPPi, min = 2/1025, i = 0, 1, 2, …7. • Once the WSTA receives new TCPP values from the HC, it reverts to the HC-coordinated contention by immediately adopting the new TCPP values.

6. Backoff Timer • A WSTA uses its latest PP to (re)set its backoff timer. • Conceptually, it generates a new pseudorandom number, X, at each idle slot (or after a busy channel becomes idle) to decide whether to transmit or not. • If X  PP, the ESTA sends a frame at the beginning of the next slot. • Operationally, it repeats the above steps within a slot time to search for the equivalent backoff time and hence sets the backoff timer. PP = P1 PP = P2 PP = P3 X  P1 BC = 0 X > P2 BC = 1 Idle Slot 1 X > P2 BC = 2 Idle Slot 2 X  P2 BC = 2 Idle Slot 1 X > P2 X > P2 BC = 1 Idle Slot 1 X > P2 BC = 2 Idle Slot 2 X > P2 BC = 3 Idle Slot 3 X > P2 BC = 4 X > P3 BC = 1 Idle Slot 1 X  P3 BC = 1 Transmits Transmits Transmits Conceptual Persistent Contention Equivalent Backoff Setting X  P1BC = 0 X > P2 BC = 1 X > P2 BC = 1 X > P3 BC = 1 X > P2 BC = 2 X > P2 BC = 2 X  P3BC = 1 X  P2BC = 2 X > P2 BC = 3 X > P2 BC = 4 Unexpired backoff timer Reset X > P2 BC = 5 BC: Backoff Counter (BC  255) Over expanded interval X > P2 BC = 6 X > P2 BC = 7 Each X is a new pseudorandom number uniformly distributed between 0 and 1 Over contracted interval X  P2BC = 7

7. 0 1 1 0 0 0 0 0 1 0 1 0 0 0 1 0 Maximal-Length Linear-Feedback Shift Register • Pseudorandom integer generation • Each shift of an m-stage maximum-length shift register produces an m-bit binary pseudorandom integer represented by the bits stored in the register. • The pseudorandom integers so generated are uniformly distributed over (0, 2m] and have a period of 2m – 1. • Pseudorandom number generation • Such pseudorandom integers divided by 2m become pseudorandom numbers uniformly distributed over (0, 1]. • Pseudorandom numbers can be generated in this way as fast as the clock frequency. • Maximum-length shift registers are also widely used in generating CRC (FCS) parity check symbols. 16-Stage Maximum-Length Linear-Feedback Shift Register (LFSR)

8. Frame Transmission • Backoff Timer Setting and Countdown (Review of External Access) • A WSTA (re)sets its backoff timer by repetitive search for a pseudorandom number, X, such that X PP. • A WSTA decrements its backoff timer just as a DCF STA does, and hence uses the same station machine as for DCF. • A WSTA transmits a frame when its backoff timer expires. • Local Selection (Internal Access) • A WSTA selects a frame for transmission from a local TC of priority k such that sum (TCPP0, …, k – 1) < X sum (TCPP0, …, k). • sum (TCPP0, …, k) = TCPP0 + TCPP1 + … + TCPPk, sum (TCPPk – 1) = 0 for k= 0, and TCPPj = 0 if TC of priority j is locally inactive. • Retry • MIB attributes of aMaxTransmitMSDULifetime,dot11ShortRetryLimit, and dot11LongRetryLimit apply to frame transmissions from individual TCs. ... ... TCPP0 TCPP1 TCPPk TCPP7 0 X PP 1

9. Internal Selection • Example TCPP0 TCPP1 TCPP2 TCPP3 0 X PP 1 FALSE FALSE TRUE FALSE TCPP0 + TCPP1+ TCPP2 < X TCPP0 + TCPP1+ TCPP2 + TCPP3 TCPP0 < X TCPP0 + TCPP1 TCPP0 + TCPP1 < X TCPP0 + TCPP1+ TCPP2 0 < X TCPP0 Queue 0 Queue 1 Queue 2 Queue 3

10. TCPP Update and Load Control • Control Criterion • Channel is optimally loaded /utilized when time on idles = time on collisions. • Time on an idle = a slot time. • Time on a collision = longest transmission time of colliding stations + Ack transmission time + SIFS + DIFS. • Control Mechanism • HC decreases TCPPs if channel is overloaded and vice versa. • Contending WSTAs immediately respond to a change in TCPPs. • Control Procedure • Compute the normalized difference, D = (TI - TC ) / T, between the time on idles, TI, and the time on collisions, TC, over the time, T, allocated to contention in the CP since the last time when a TCPP update was broadcast. • When D  D0 or when a beacon is to be sent, update TCPPs as follows: • TCPP0 TCPP0 + G  D, and TCPPk = Ck TCPP7, k =1, 2, …, 7, where D0 and Ck are preset numbers, and G is positive. • Algorithm self-stabilizing: TCPPk , TC , D , TCPPk , and vice versa.

11. Fairness and Differentiation • Fairness • All TCs of equal priority transmit with the same TCPP. • Anytime -- before or after collision. • Anywhere -- at the same WSTA or at different WSTAs. • Differentiation • Relative differentiation: Higher priority TCs contend with larger TCPPs. • TC Access probability ~ TCPP. • Lower priority TCs not starved. • Absolute differentiation: Some TCs may be stayed from contention. • TCPPs for stayed TCs set to 0. • Higher priority TCs not impacted by lower priority TCs. • Minimum bandwidth guaranteed for selected TCs. • Maximum bandwidth imposed on certain TCs. • Collision avoidance enhanced. • Collision resolution accelerated. • Some stayed TCs served by contention-free access for better QoS support.

12. Implementation Complexity • P-DCF (CSMA with Adaptive Contention) • No new IFS rules. • Single backoff timer per WSTA. • No internal conflicts. • V-DCF (Stacked DCF) • New IFS rules. • Multiple backoff timers per WSTA. • Internal conflicts. • Extra logic is required to detect and resolve internal collision.

13. Backoff, Collision, and Delay • Backoff of Smaller Variation • Backoff of Larger Variation • Reducing Collision  Reducing Access Delay/Jitter Collision Success Backoff Time Access Delay Success Backoff Time Access Delay

14. Backoff Delay/Variation versus Access Delay/Variation • Backoff of Smaller Variations • Backoff of Larger Variations Collision involving frame 1 Transmission of another frame Collision involving frame 1 Transmission of another frame Collision of other frames Transmission of frame 1 only Channel busy Backoff Time Access Delay 1 Backoff Jitter Collision involving frame 2 Transmission of another frame Transmission of frame 2 only Channel busy Backoff Time Access Delay 2 Access Delay Jitter Delay Jitter > Backoff Jitter Access delay, and hence access delay variation, of a frame depends not only on backoff time and backoff variation but also on many other factors. Reducing collision (which causes long delay) is more effective in minimizing access delay and variation than reducing backoff delay or variation. A smaller backoff variation does not lead to a smaller access variation. A CSMA protocol with constant backoff (and hence zero backoff variation) is most likely to yield larger access delay and variation than a CSMA protocol with uniform backoff (and hence nonzero backoff variation). Collision involving frame 1 Transmission of another frame Transmission of frame 1 only Channel busy Backoff Time Access Delay 1 Backoff Jitter Transmission of another frame Transmission of frame 2 only Channel busy Backoff Time Access Delay 2 Access Delay Jitter Delay Jitter < Backoff Jitter

15. Access Delay and Variation (Simulation Result) Operation “recovered” from congestion as frames exceeding retry counts were discarded.

16. Instantaneous Delays for 4-Priority P-DCF TCPP update per 20 ms interval

17. Average Delays for 4-Priority P-DCF TCPP update per 20 ms interval

18. Instantaneous Delays for 4-Priority V-DCF

19. Average Delays for 4-Priority V-DCF

20. P-DCF Performance Improvement • IEEE journals report adaptive contention (P-DCF) to achieve more than 30% throughput than binary exponential backoff. • PP. 146-149 in F. Cali, et al., “Dynamic Tuning of the IEEE 802.11 Protocol to Achieve a Theoretical Throughput Limit,” IEEE INFOCOM’98. • P. 1783 in . F. Cali, et al., “IEEE 802.11 Protocol: Design and Performance Evaluation of an Adaptive Backoff Mechanism,” IEEE J. Select. Areas Commun., vol. 8, Setp. 2000. • Our own simulations also show P-DCF to have significant throughput, delay, and jitter improvement over binary exponential backoff. • Adaptive contention is robust to PP miscalculations. • 20 - 50 ms is adequate for TCPP updates. • Performance improvement becomes even more substantial in high population areas such as in enterprise environments. • Binary exponential backoff begins to fail as user population increases.

21. Differentiation by CWmin? • CWmin = 31 slots for 802.11b • This appears to be a choice of the right tradeoff between minimizing idles and collisions for a CSMA based on binary exponential backoff rules. • Using smaller CWmin results in increased collisions while using larger CWmin leads to excessive idles for typical channel loads. • Changing CWmin values without changing other CWs leads to worse throughput and delay performance for all priorities than legacy DCF. • CWmin = 15 slots for 802.11a • Choosing CWmin of 7 and 3 for higher-priority TCs leads to intensive collision and diminished throughput. • Using CWmin of 31 and 63 for lower-priority TCs yields QoS service much worse than simply following legacy DCF. • There is no much room for changing CWmin values under binary exponential backoff rules.

22. Adaptation for CWmin? The leftover backoff times of various TCs are inestimable to EAP/HC at the time of setting new CWmins. Setting CWmins without knowing leftover backoff times causes new frames to collide with backoff frames. Each collision costs much channel time, aggravates congestion state, and results in more collisions in the future. Doubling CW for colliding TCs at low load unnecessarily delays frame transmission and decreases channel throughput. Doubling CW for colliding TCs does not necessarily alleviate congestion state --collisions formed from past backoffs will still occur. Doubling CW for colliding TCs over-penalizes the colliding TCs, even more so for those that had collided before. … … … … … … … … … …   Randomizing backoff times without drastically increasing CW at low load is adequate and improves delay and throughput performance. CWmin update for new frame arrivals only Resetting CWmins cannot stop collisions developed in the past and bound to occur in the future. Having multiple backoff timers per station makes delay and throughput ever more sensitive to CWmin values. Doubling CW for colliding TCs substantially downgrades the access priority for both retried and new frames of those TCs, compared to frames of TCs not undergoing collision resolution. Contention-based and contention-free transmissions

23. Differentiation by IFS? • Just not desirable. • Lower-priority traffic suffers excessive delays. • Additional slots in expanded IFS wastes channel bandwidth. • Especially so with HCF. • HC can do a much better job with tier access and absolute priority. • WSTAs using differing IFSs for prioritized access complicate their operation as well as HC’s.

24. Concluding Remarks • Binary exponential backoff has been considerably criticized for its poor throughput and delay performance inside and outside 802. • Adaptive contention has been investigated by several experts: • L. Kleinrock, inventor of the Internet technology. • F. Tobagi, author of CSMA and its various variants. • R. Gallager, communications and networking authority. • Performance improvement from CSMA with adaptive contention (P-DCF) has been shown to be significant in the IEEE literature. • Increasing bandwidth demand warrants such an improvement. • QoS is better supported by an efficient protocol. • Lessons have been learned from Ethernet and are worth learning. • CSMA/CD degrades rapidly in performance as node population increases. • CSMA without collision detection costs much more collision bandwidth, and hence performs even worse, than CSMA/CD. • Wireline Ethernet needs to become faster and faster. • Wireless LANs have scarce spectrum resources. • IC technology is much more advanced and affordable than 1970’s when Ethernet was first developed.

25. Sample References Books: 1. D. Bertsekas and R. Gallager, Data Networks, 2nd ed., Prentice Hall, NJ, 1992, Chapter 4. 2. A. S. Tanenbaum, Computer Networks, 3rd ed., Prentice Hall, NJ, 1996, Chapter 4. Papers: 1. L. Kleinrock and/or F. Tobagi's CSMA and CSMA/CD papers published between 1975-1985. 2. F. Cali, et al., “IEEE 802.11 Protocol: Design and Performance Evaluation of an Adaptive Backoff Mechanism,” IEEE J. Select. Areas Commun., vol. 8, Setp. 2000, pp. 1774-1786. 3. F. Cali, et al., “Dynamic Tuning of the IEEE 802.11 Protocol to Achieve a Theoretical Throughput Limit,” IEEE INFOCOM’98, pp. 142-149.

26. Questions Available Technology = Enhancement ? Degraded DCF = Enhanced DCF ?? Stacked DCF = Simplicity = QoS ???