190 likes | 284 Views
Cancellation of aggregate Multicast feedback – measurement results. Date: 2010-07-09. Authors:. Abstract.
E N D
Cancellation of aggregate Multicast feedback – measurement results Date: 2010-07-09 Authors: Jochen Miroll
Abstract • This presentation is an update on the Leader-based aggregate feedback Protocol (LBP) proposal previously made to TGaa by the authors and provides measurement results obtained on a consumer 802.111 hardware test bed • The feedback cancellation probability in the worst case of LBP is measured and compared to previous theoretical / simulation results • These results have also been published and presented at the IEEE ISCE 2010 conference in June 2010 Jochen Miroll
Motivation • 11aa is standardizing Multicast ARQ: MRG • Gathering per-receiver feedback, the overhead due to the positive ACKs grows linearly with the number n of receivers • How does 11aa MRG compensate for this increased overhead? • Aggregation of multiple frames: single-TID, uncompressed Block-ACK (802.11n) for MRG • Per-frame ACK becomes multi-frame Block-ACK bitmap for the last k frames • Still: overhead increases linearly with receivers n • How to get rid of the dependency on n? • We have previously proposed a leader-based Multicast retransmission scheme to 11aa Jochen Miroll
NACK NACK ACK Feedback aggregation in the same time slot • All receivers provide feedback, butit is aggregated in the same time slot - then(n = number of receivers) • overhead(n) = overhead(1) • Idea: Introduce NACK • Transmit a data frame • Then, ask for ACK/NACK • If STA i has received the frame:it responds with an ACK • If STA j did not:it responds with NACKat the same time ? AP1 STA 1 STA 2 STA 3 STA 4 Jochen Miroll
Feedback cancellation premise • If ACK and NACK are approx. equally strong • Is it possible to cancel an ACK by a simultaneous NACK and thus enforce a retransmission? • The „capture effect“: • Describes the phenomenon that a frame (e.g. ACK) may be received correctly in the presence of another, similarly strong (e.g. NACK) • Main reasons for „imperfect collision“ • Locking the preamble and then Viterbi decoding the locked-onto frame is a very robust mechanism. • E.g.: ACK is BPSK, rate ½ and only 14 Bytes in length.It is the most robust 802.11 frame(OFDM: few dB difference between ACK and NACK may suffice to „capture“) Jochen Miroll
Earlier comments from TGaa (resolved) • Will feedbackcancellationactuallywork? • answer: Yes, collisions are happening all of the time • answer: No, due to the capture effect • We have consequently provided Matlab and ns-2 results for feedback cancellation to Tgaa • cf. doc.: IEEE 802.11-09/1150r2 • Provided in this document: measurement results using real and cheap 802.11 hardware Jochen Miroll
Leader-based feedback cancellation • Idea: Multicast is essentially handled as a unicast connection to a „leader receiver“ • „Non-leaders“ transmit a NACK if a frame is lost • Target: Larger Multicast groups (large n) • If ACK survives the somewhat weaker NACK, does it survive many? Does it survive many equally strong, many somewhat stronger? • Intuitive leader selection: choose the „weakest“ receiver (as seen by the AP, no power control, just due to path loss) • If no loss: Leader’s ACKs can be received (ACKs are most robust) • Else: Expect a good chance that whenever several somewhat stronger NACKs are transmitted at the same time, the Leader’s ACK will be cancelled Jochen Miroll
Aggregation through Leader-Based feedback cancellation Protocol (LBP)cf. doc.: IEEE 802.11-09/0290r1 optional SEQ# indicator and NAV updater to synchronize aggregate feedback Jochen Miroll
Feedback cancellation constraint • Failure of feedback cancellation results in uncorrectable packet loss at non-leaders • (i.e. capture of ACK happens, no collision) • Question that arises: • What is the error floor in the worst case? • What is the worst case for the leader-based feedback cancellation approach? • Intuitively: the „weakest“ receiver can not be distinguished • All receivers on average experience the same SNR • We assume that all are sending approx. equally strong feedback Jochen Miroll
Feedback cancellation measurements • Examine two different cases of how feedback aggregation may be implemented • In the WLAN card‘s real-time OS • In the WLAN card‘s host OS (e.g. Linux) • Implications • Cards allow for strict timing constraints (similar to 802.11 ACK, ±900ns), so we can examine short feedback • Host OS is less accurate in timing, thus we examine feedback cancellation with frames of several tens of Bytes Jochen Miroll
Feedback cancellation test setup (1) • We have used real consumer 802.11 hardware • Limited freedom in implementing MAC algorithms • But: We can fix some parameters in cancellation experiments • Here: Non-leaders transmit different frames • Examine different frame sizes and timings with what is possible… …out of the box: Let positive feedback be a 6 Mbps ACK and the negative feedback be a 12 Mbps ACK …own implementation: Driver level ACK/NACK implementation Jochen Miroll
Feedback cancellation test setup (2) SEQ frametriggersfeedback, assumethisisthequestion „didyougetthedataframe“ Jochen Miroll
Non-leader 1 Non-leader n-1 AP Leader Feedback cancellation test setup (3) • To obtain independence from the (fading) environment: • Move receivers slowly around the AP, changing their positions in the environment • Periodically change the roles (leader, non-leader) of the receivers(always have exactly 1 leader) Jochen Miroll
CDF of SNR at receivers is very steep ~identical channel conditions for all receivers on average Error free reception rates of different frames at the end of measurement run yield valid results SEQ (trigger) loss?loss rate < 0.1% Validation of test setup Jochen Miroll
Test results (representative example) • 1 leader, 3 non-leaders • Why? An example, assume • But: Assume large n • Virtuallyno SEQ loss • ~89% feedbackcancellationsuccessprobability • Resultseemsindependentofframelengthandtiming • Worstcaseresults (whereleader-selectionwould not work) Jochen Miroll
Theoretical / Simulation results • Compare with ns-2 results • Scenario: Rayleigh fading channel, equal AP-STAs distance • feedback cancellation rate is about 76% for 2, • more than 90% for more than 2, and • already 99% for 5 receivers • Again: worst case whereleader selection fails Jochen Miroll
Conclusion • Scalable Multicast error correction can be achieved by aggregation through cancellation • Real test bed results are backed up by simulations • Channel will not be arbitrarily reliable but limited by an error floor • Combined MAC-layer and “Application Layer” error correction feasible • Assume overlay packet erasure FEC • Audio/Visual streams typically can tolerate errors • Residual error requirement can be dealt with on layers above MAC Jochen Miroll
Questions? (a further presentation will propose how this scheme should be incorporated into 11a) Jochen Miroll
Recap: Hybrid LBP (HLBP)*cf. doc.: IEEE 802.11-09/0290r1 Phase I Transmit a block of frames, as in MRG BA. Here: systematic FEC part Phase II Parity phase. Instead of BAR/BA, do AggregateAckRequest/AggregateAck * Assume e.g. DVB-IPDC or Raptor code on upper layer, MAC somehow knows which packets are systematic (DATA) or parity Jochen Miroll