Re Real-tim time e Dis istr trib ibuted ed MIM IMO Sy Systems Hariharan Rahul Ezzeldin Hamed, Mohammed A. Abdelghany, Dina Katabi
De Dense W Wir irele less Ne Networks • Stadiums • Concerts • Airports • Malls
In Interf erfer eren ence L e Limits Wi ts Wirel eless T ess Throu oughput APs cannot transmit at the same time, in the same frequency à Take turns to avoid collisions Ethernet Et … AP AP3 AP N AP AP2 AP AP1 AP … User 3 Us User N Us User 2 Us User 1 Us Total Wireless Throughput Stays Constant à Each AP gets 1/N of the total throughput
Dis Distrib ibuted M MIMO is is t the H Holy ly Gr Grail ail Distributed protocol for APs to act as a huge MIMO transmitter with sum of antennas Ethernet Et … AP AP3 AP AP N AP2 AP AP1 AP … User 3 Us User N Us User 2 Us User 1 Us N APs à N times higher throughput
Mu Much recent work rk in mo moving distri ributed MI MIMO MO fr from theory y to practice Ho However er, we e still till do do no not t ha have e real eal-tim time e dis istr trib ibuted ed MI MIMO MO systems ms operating on independent de devi vices es wi with h thei heir own wn clocks!
Wh Why y aren’t n’t we the here yet? • Distributed MIMO needs accurate channel estimation à High overhead process that could eat up all the gains. • Need distributed power control. • Need an architecture that can support these complex operations in real-time.
Wh Why y aren’t n’t we the here yet? • Distributed MIMO needs accurate channel estimation à High overhead process that could eat up all the gains. • Need distributed power control. • Need an architecture that can support these complex operations in real-time.
Ch Channel Es Estima mation and Feedback Et Ethernet … AP AP3 AP N AP AP2 AP AP AP1 … User 3 Us Us User N Us User 2 User 1 Us N Channel Estimation Packets N 2 Channel Measurements Need to do this periodically as environment changes
Ch Channel Feedback k Ov Overh rhead
Re Reciprocity in Traditional MIMO ℎ &'(),% ℎ "#,% Client Access Point ℎ "#,* ℎ &'(),* • Reciprocity is the property that the ratio of downlink channels is equal to the ratio of uplink channels up to a constant. • This constant is the ratio between hardware chains of AP antennas. • Allows us to estimate this constant once and use it for all future uplink transmissions and across clients.
Wha What ha happe ppens ns wi with h Distribut buted d MIMO? O? AP[1] AP[2] Separate devices à Different Crystals à RF chains have oscillator offset relative to each other
Tr Traditional Reciprocity does not work with Dis Distrib ibuted M MIMO The “constant” is no longer constant, but changes rapidly with time. Theorem: The downlink and uplink channel ratios can be written as: 8 >?,6 8 9:;<,6 8 9:;<,= = 𝐷 * 𝑢 × 8 >?,= where 𝐷 * (𝑢) = 𝐷 * (0)×𝑓 3*∆5 6 7
Re Reciprocity and Distributed MIMO Calibration Calibration Parameter is rapidly time varying à Cannot do one-time calibration Need to repeatedly calibrate: • for uplink transmissions from every client • at every AP
Me MegaMI MIMO MO 2. 2.0 0 Ca Calibration for r Reciprocity • Avoids the overhead of repeated calibration • Distributed mechanism for updating calibration parameters at slaves with no overhead
Me MegaMI MIMO MO 2. 2.0 0 Ca Calibration Formu rmulation Master AP Slave AP 𝐷 * (𝑢) = 𝐷 * (0)×𝑓 3*∆5 6 7 • Compute the initial calibration parameter, 𝐷 * (0) • Update the calibration parameter at time t by estimating 𝑓 3*∆5 6 7
Me MegaMI MIMO MO 2. 2.0 0 Initial Ca Calibration Master AP ℎ * ℎ % Slave AP 1. Measure channel ℎ % from Master AP to Slave AP 2. Measure channel ℎ * from Slave AP to Master AP 𝐷 * 0 = ℎ * 3. Compute Initial Calibration Parameter 𝐷 * 0 as ℎ % 4. At slave, store 𝐷 * 0 and ℎ % as ℎ % (0)
Me MegaMI MIMO MO 2. 2.0 0 Ca Calibration Update 𝑢 𝐵𝑑𝑙 Client Master AP Packet ℎ % (𝑢) Slave AP 1. Client transmits packet à Master and Slave measure uplink channels from client 2. Master sends sync trailer (Can leverage Wi-Fi ack) Slave measures channel ℎ % 𝑢 from master. ℎ % 𝑢 = ℎ % 0 ×𝑓 3∆5 6 7 3. * ℎ % (𝑢) Recall that each slave has ℎ % 0 . Each slave computes 𝑓 3*∆5 6 7 = 4. ℎ % (0) Consistent channel estimates using reciprocity at 𝐷 * (𝑢) = 𝐷 * (0)×𝑓 3*∆5 6 7 5. Each slave computes the updated calibration parameter all APs 6. Each slave computes the corrected downlink channel using the updated calibration parameter
Me MegaMI MIMO MO 2. 2.0 0 Procedure • Preparing Calibration Constants • Master AP transmits a reference packet • All slaves follow with a response • Each slave calculates its calibration parameter • Channel Estimation • Performed for each uplink transmission from a client • The master AP follows with an ACK (Sync trailer) • Each slave calculates its downlink channel using the corrected calibration parameter • Joint Transmission • The same as MegaMIMO 1.0
Wh Why y aren’t n’t we the here yet? • Distributed MIMO needs accurate channel estimation à High overhead process that could eat up all the gains. • Need distributed power control. • Need an architecture that can support these complex operations in real-time.
Th The Need for Automatic Gain Control (AGC) RF Chain ADC Digital Processing +1 V Converts analog Works in analog Decodes digital samples signal to digital domain -1 V samples • ADC accepts signals in a specific range • RF chain converts received signal to ADC range • AGC is an adaptive algorithm to perform this conversion
AGC in Traditional MIMO AG AP applies the same gain to all receive antennas h 14 h 11 h 12 h 13 14 11 h 24 h 21 h 22 h 23 h 31 h 32 h 33 h 34 h 44 h 41 h 42 h 43
AG AGC in Traditional MIMO AP applies the same gain to all receive antennas 𝜷 𝜷 h 14 𝜷 h 11 𝜷 h 12 h 13 14 11 h 24 h 21 𝜷 h 22 𝜷 h 23 𝜷 𝜷 𝜷 𝜷 𝜷 𝜷 h 31 h 32 h 33 h 34 𝜷 𝜷 𝜷 𝜷 h 44 h 41 h 42 h 43
AG AGC in Distributed MIMO Each AP-client link has an independent gain 𝜷 12 𝜷 13 𝜷 14 h 14 h 11 h 12 h 13 𝜷 11 12 13 14 14 11 11 𝜷 22 𝜷 23 𝜷 24 h 24 h 21 h 22 h 23 𝜷 21 22 23 24 21 h 31 𝜷 32 h 32 𝜷 33 h 33 𝜷 34 h 34 𝜷 31 32 33 34 31 𝜷 42 𝜷 43 𝜷 44 h 44 𝜷 41 h 41 h 42 h 43 42 43 44 41 We need a protocol for ensuring that the multipliers are the same despite being applied on different boxes
Co Comp mpensa sating for r the AGC • AGC typically has a coarse power setting à Need to convert to a complex 𝛽 value. • This conversion is not known a priori . • MegaMIMO 2.0 learns this conversion factor. • Each antenna transmits a signal. • Receiver sets gain to a particular coarse value, and measures received channel • Repeats across all coarse gain settings • Needs to be recalibrated infrequently to account for drift of analog components.
Wh Why y aren’t n’t we the here yet? • Distributed MIMO needs accurate channel estimation à High overhead process that could eat up all the gains. • Need distributed power control. • Need an architecture that can support these complex operations in real-time.
Me MegaMI MIMO MO 2. 2.0 0 PHY-MA MAC C Architecture • 802.11 PHY is a complex system: power adaptation, rate adaptation, encoding and decoding at various modulations and code rates etc. • Traditional PHY layers only have local control and coordination with an on-board MAC. • Distributed MIMO requires distributed control and coordination across multiple transmit and receive chains. • We design an architecture that provides hooks to/from the PHY to enable this distributed control efficiently in hardware.
Pe Performance
Im Implem plemen entatio tion • Implemented on Zed Board and FMCOMMS2 RF Front End • PHY and real time MAC implemented on Zynq FPGA • Control Plane implemented on embedded ARM core
Evaluation Ev • Indoor Testbed simulating a conference room • 4 APs transmitting to 4 clients • Line of sight and non line of sight scenarios • Mobility • Environment • Users • Metrics • SNR obtained by users during joint transmission • Total throughput
Re Reciprocity vs. Feedback
Re Reciprocity vs. Feedback
Re Reciprocity vs. Feedback Re Reciprocity matches feedback across the range of SNRs à Ca Calibration is accurate
Me MegaMI MIMO MO 2. 2.0 0 vs. Traditional 802. 802.11 11
Me MegaMI MIMO MO 2. 2.0 0 vs. Traditional 802. 802.11 11
Me MegaMI MIMO MO 2. 2.0 0 vs. Traditional 802. 802.11 11 3.3-3.6x Me MegaMI MIMO MO 2. 2.0 0 with h reci ecipr proci city pr provides des the he expect pected ed scaling ng gains ns acr cross the he ra range of SNRs
Reciprocity Th Throughput Gain with Mobility Environmental Movement Client Mobility
Reciprocity Th Throughput Gain with Mobility No single feedback interval is optimal across all scenarios. Environmental Movement Client Mobility Reciprocity outperforms explicit feedback.
Recommend
More recommend
