Response and Rebuttal To Doc. 01/011

1. Response and RebuttalTo Doc. 01/011 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), (919) 463-1043 (Schrum), (214) 480-6908 (Turki) (214) 480-4349 (Shaver), (214) 480-2344 (Shoemake) jinmengho@ti.com, sschrum@ti.com, khaled@ti.com shaver@ti.com, shoemake@ti.com

2. Introduction • Results of analysis and simulation contradict with the direct comparison between p-DCF and v-DCF appearing in doc. IEEE 01/011, entitled “Comparing V-DCF with other EDCF Proposals”, authored by Wim Diepstraten, et al. • Some comments and clarifications on the various statements set forth in that document are in order. • Remarks quoted from that document appear in italic shadow font in this document. • Replies by us to the quoted remarks are interspersed and appear in regular font.

3. Slides 3 and 4 of doc. 01/011 • Target functionality is: • We need a distributed mechanism that can stabilize the throughput for high contention situations, preventing collision avelange effects when the load and number of contenders increase. • The new retry Backoff procedure (for v-DCF): • Start exponential backoff by CW doubling after the first retry, on all Q’s, so per station. • If CW is not doubled for the first retry, contention is not reduced. This seems not to be a mechanism that can stabilize the throughput for high contention situations, but rather to induce and accelerate collision avalanche effects when the load and number of contenders increase. Simulations have supported this conjecture. • If traffic in the next Q is likely going to the same destination then the CW doubling should apply per station, else it is more effective to apply CW doubling only for the Q experiencing the Retry (AP). • There would not be normative station behavior in this way.

4. Slide 6 of doc. 01/011 • Effect of different probability distributions: • V-DCF uses a uniform distributed backoff mechanism • resulting in a low delay variance • If low delay variance of backoff were the objective of the MAC protocol design, we would have an easy task -- by simply setting the backoff to a constant or even to zero, rather than by drawing the backoff from a probability distribution. • and gives higher priority to “older” frames which have already counted down part of their backoff. • Recall that CW doubling for colliding frames is a severe penalty for “older” frames. Also see comments elsewhere. • Allows for immediate access if CCA>DIFS, which is advantageous for high priority frames which Q’s are not backlogged. • Immediate access by no way guarantees immediate success, especially after the medium has been busy for a while--immediate access may lead to immediate collision, which is bad for every frame.

5. Slide 6 of doc. 01/011 • P-DCF uses a geometric access distribution mechanism • resulting in high delay variance (which is very undesirable for QoS) • Given the same mean delay of the first backoff, p-DCF has a higher delay variance of the first backoff than v-DCF. However, this does not mean p-DCF will have a higher delay variance of the access delay than v-DCF, where the access delay accounts for all the backoff times (including those for retransmissions) in successfully sending a frame from a transmitter to a receiver, and it is this access delay, but not the individual backoff delay, that is of more concern to QoS. If fact, simulation results show p-DCF, because of its adaptability in coordinating contention, to produce much smaller access delays than v-DCF. • Zero backoff CSMA would give zero delay and zero variance of the first backoff, but would it perform better than binary exponential backoff CSMA? Non-slotted Aloha even gives zero delay and variance of the wait time for the first transmission, does it provide better QoS than binary exponential backoff CSMA? It is much more important to minimize the number of unsuccessful transmissions (and hence the delay and variance of the access delay), but not the delay and variance of the backoff time for the first transmission.

6. Slide 6 of doc. 01/011 • P-DCF uses a geometric access distribution mechanism • is memoryless, so does not favor frames that are already backing off for some time. • It is precisely the memoryless property that does not disfavor the frames that are already in backoff for some time. Several people, including Bob Meier of Cisco, on the January 3 conference call clearly explained this issue which was discussed on slide 15 of doc. 00/467 (slide 12 of doc. 00/467r1). • Frequent updates does make this worse. • Again, the memoryless property makes the expiration time of an unexpiring backoff timer statistically unchanged after an update with the same permission probability. An update with a larger permission probability for the case of light load rather foreshortens the expiration time, while an update with a smaller permission probability for the case of heavy load prolongs the expiration time. This is the advantage of update and hence adaptation. • Does always go through backoff (no immediate access). • See comments on slide 4 relating to v-DCF. With p-DCF, when the load is light, the permission probability is large and hence the backoff is very small. When the load is heavy, the permission probability is small, and hence the backoff is relatively large. Such adaptation optimizes channel throughput for both low and high loads.

7. Slide 7 of doc. 01/011 • Stability control mechanisms: • V-DCF uses an exponential backoff behavior after the first retry. • Its stability control mechanism is fully distributed, and does not depend on a centralized congestion control mechanism. • V-DCF does require a centralized control mechanism to provide CWmin’s. The simulation results indicate that if CW is chosen incorrectly from the onset, the channel quickly becomes unstable. On the other hand, the central control algorithm we developed for p-DCF (see slide 13 of doc. 00/467, or slide 14 of doc. 00/467r1) is self-stabilizing--if the permission probability values are overestimated or underestimated, they will be automatically adjusted back at the next update. Moreover, backup rules are defined that enable wireless ESTAs, when not hearing a contention control element in the beacon or the beacon itself, to choose permission probabilities on its own in a way that are completely comparable to the distributed CSMA with backoff.

8. Slide 7 of doc. 01/011 • Stability control mechanisms: • V-DCF uses an exponential backoff behavior after the first retry. • It prevents for collision avalanches due to congestion. • User experience from Ethernet, and results from simulation conducted by a large number of researchers and developers around the world, indicate that binary exponential backoff quickly leads to collision avalanches as congestion develops. • And is targeted to avoid effect of hidden nodes and overlap interference, by avoiding overlap with the hidden message. • V-DCF is essentially multiple DCFs running at the same ESTA for multiple local TCs, with CWmin’s determined and broadcast by a central controling entity. Thus, it does not cope with hidden nodes and overlap interference better than the legacy DCF. Hidden nodes can prevent the central controlling entity’s CWmin feedback from coming or reaching the contending stations, and hence can cause the v-DCF to malfunction. • Relates to interference situation at the receiver not visible to the transmitter • ?

9. Slide 7 of doc. 01/011 • Stability control mechanisms: • P-DCF stability control mechanisms are not clear. • An unambiguous control algorithm is provided on slide 13 of doc. 00/467 (slide 14 of doc. 00/467r1) which is self-stabilizing as explained there (see also slide 7 of this doc). Menzo, a co-author of v-DCF, appears to be advocating the use of this algorithm for v-DCF, although this was developed for adaptive contention. • Either an autonomous decrease of the PPC based on retry event. • Yes, the rules for decreasing the TCPP values are provided for cases where the central control information is not being received. • Or a centralized control mechanism to control the PPC. • If this is the main mechanism then it does fully depend on a centralized entity for stability. • See the comments above. • Conclusion: A stability control mechanism is needed that is fully distributed and does not depend on other stations. • V-DCF depends on a central controlling entity to determine CWmin, which, if not chosen to reflect the channel load, makes the channel unstable quickly. Again, such behavior is evidenced from simulation results.

10. Slide 8 of doc. 01/011 • V-DCF: • V-DCF is very similar to the legacy DCF approach. • This leads to similar performance of the legacy DCF. • Can be implemented as n*DCF (parallel DCFs). • In this case the complexity is n times the legacy DCF. • Or a Scheduler function combined with a “Delta backoff DCF” as presented during the Tampa meeting, and described in doc 00/399 • In this case the scheduler has to constantly interact with the “Delta backoff DCF” and is complex and potentially power consuming. • This scheduler is also effective to select candidate for CF-polled TxOp. • This will create very undesirable effects: Frames, especially from periodic sources like video and voice, will be also sent by periodic and hence intensive contention--which greatly reduces channel throughput. • Minimal complexity increase compared to plain DCF. • ??? • Roughly the same number of computations per frame needed as in legacy DCF. • ???

11. Slide 8 of doc. 01/011 • P-DCF: • In its simplest form it requires a random number generation per slot • p-DCF can operate by persistent contention or adaptive backoff as shown in Doc. 00/467 (or Doc. 00/467r1). If it operates by persistent contention, it transmits with a probability following an idle slot, but not on every slot. Random numbers needed on idle slots can be generated in advance (in non-real time) and stored in a very small cache, which can be replenished when the station is idle. If it operates by adaptive backoff, its complexity is the same as the legacy DCF except for the calculation of the backoff in terms of a log ratio (see slide 18 of Doc. 00/467, or slide 19 of Doc. 00/467r1). Again this log ratio may be calculated and stored in advance. Bob Meier suggested such a scheme. • This is a major computational increase for FW based solutions. • This is not a major computational increase. Besides, CC/RR, which is already adopted in the baseline, also operates by probabilistic contention. • Conclusion: V-DCF approach is less computation intensive. • We do not agree.

12. Slide 9 of doc. 01/011 • What is the motivation that justifies the change to P-DCF? • Better QoS support and more efficient channel utilization. • Properties that are claimed to be important: • Better analyses possible (without exponential backoff) • This is correct, but for practical benefits--better estimates of channel load are possible. • Memoryless so can be better analyzed • The memoryless property is not there so that p-DCF can be better analyzed. It is there as a desirable consequence of the p-DCF operation--it allows to reset, and hence, adapt, the backoff timer in response to load changes without placing backoff stations at a disavantageous position relative to stations not in backoff. See also comments on slide 6 of this document, or discussions on slide 12 of doc. 00/467r1. • Faster update from one parameter set to another allows better optimization control • A clarification: Faster update with p-DCF does not necessarily mean that the contention parameters need be updated by a central controlling entity more frequently than with v-DCF. Rather, v-DCF ESTAs, when having a new frame to send, choose a backoff time based on the latest broadcast CWmin’s from the central controlling entity, and vary the CW, and hence perform the contention, on their own thereafter, regardless of subsequent contention parameter updates from the central controlling entity. They will not update their CWmin’s until the next new frame is ready to be sent. In contrast, p-DCF ESTAs adjust their backoff times with each contention parameter update from the central controlling entity.

13. Slide 9 of doc. 01/011 • What is the motivation that justifies the change to P-DCF? • Properties that are claimed to be important: • No internal collisions • This assures that multiple TCs at the same ESTA do not collide, and hence be penalized, more frequently than single TCs at other ESTAs, thereby all TCs of equal priority, regardless of where they are located, are always contending with equal probability. Moreover, no additional steps are needed in resolving internal collisions. • Better performance due to better control. • Yes. • How different are they and how important are these factors? • They greatly improve throughput, delay, and stability as demonstrated by extensive theoretical analyses and simulation studies conducted and published by world wide experts, and by ourselves. See doc. 467r1 and the references cited therein.

14. Slide 9 of doc. 01/011 • How different are they and how important are these factors? • It is our opinion that these factors are insignificant, while controllability of both approaches are equivalent. • Both theory and simulation contradict with this opinion. The controllability of the two approaches are not equivalent as explained on slide 12 of this doc. p-DCF gives full controllability over the ESTAs’ contention, while v-DCF by design allows the control over the CWmin’s, but not CWs by which stations contend once they chose a CWmin. • A big negative for p-DCF is the huge delay variation properties of the geometric distribution. • As explained on slide 5, this statement is a fundamental misconception of the delay and delay variation in channel access. p-DCF optimizes the channel throughput systematically, and produces much smaller delay and delay variation in channel access than v-DCF as observed from the simulation results.

15. Slide 10 of doc. 01/011 • The following figures show the difference in delay jitter performance between the normal and geometric distribution? • These figures are for the backoff time, but not for the access delay. Setting the backoff time to 0 would have 0 delay jitter--better than obtained any distribution. • Clearly this is not desirable for QoS. • QoS concerns the overall access delay. • CO and CW provide separation of functions • CO creates differentiation between priorities • CW provides randomness to account for presence of other contenders • Note that the CO parameter does allow the CW to be relative small to achieve the same average access delay, which does significantly reduce the delay variance. • Using CW on binary boundaries has implementation advantages