AIMI: AP-Side only MIMO using off-the
Transcription
AIMI: AP-Side only MIMO using off-the
AIMI: AP-Side only MIMO using off-theshelf CSMA Client Radios Sungro Yoon, Bangchul Jung, Kyunghan Lee and Injong Rhee North Carolina State University 1 Performance Bottleneck Problem Collision Collision Collision Collision A base sta(on becomes the performance bo0leneck because of uplink packet collisions Collision Collision 2 Sources of Packet Collisions Hidden Terminals High Conten8on 3 Throughput Degradation in AP Networks § Throughput performance in infra-‐structured WLANs (Rohan et al., NSDI 2009) Maximum 10% degradation in ad hoc networks when n à 10 [Bianchi, JSAC 2000] 69% degradation in AP networks when n à 10 [Rohan et al., NSDI 2009] 4 Goal § To improve the network performance by adding more hardware (more antennas) and complexity … High Performance AP … High Performance AP Collision 5 Wireless Communication Primer § Mul8path Fading − Signals go through diverse paths un(l they reach to the final des(na(ons − Channel coefficients should be accurately measured at the receiver H=2 x = -‐1 Received signal y = -2 6 MIMO (Multiple Input Multiple Output) Frame 1 Frame 2 y = h x +h x +w y = h x +h x +w − − − − 1 11 1 2 21 1 12 22 2 2 1 2 Y’s are the received values on antennas X’s are the unknown transmiBed symbols. H’s are the known channel coefficients W’s are the known noise. x = Fy Find a filter F that can produce x. 7 Information Theory for MIMO Shannon’48 C= log(1+ SNR) Telatar’95, Foschini and Gans’98, Bolcskei et al’02 C ! M (1+ SNR) where M is the number of antennas. 8 Multi-‐User (MU) MIMO − Single-‐User MIMO ü One user sends mul(ple packets to the AP (or vice versa) at a (me ü E.g., IEEE 802.11n − Mul(-‐user MIMO ü Uplink: many users can transmit to the sink or AP simultaneously. ü Downlink: AP or sink can send mul(ple frames to mul(ple users simultaneously. Single User MIMO Multi-User MIMO 9 MU MIMO theory into reality § Theory focuses on coding, channel modeling § General assump8ons in theory − Nodes communicate in perfectly controlled environments OR − Cri8cal parameters (channel coefficients, CFO…) are readily given. § But… in reality, devils in detail: − Coordina(on, feedback delays, bursty traffic, hidden terminals, collisions, etc. Scheduling Power Control Is the prac8cal Practical Implementation still very challenging implementa8on possible in an unmanaged, random network anyway? Capacity Region Coding … Channel Modeling 10 Practical implementation § Yes, it is possible, but…. Coordination requires changes both in receivers and transmitters. 11 Important Practical Issues § It is not op8mality…. Deployability § Many diverse User Devices and Interfaces − Perhaps several 10s chip and device manufacturers. § They are all widely deployed − IEEE 802.11a,b,g and n. 12 Practical Deployability means .. § You can’t modify user devices − Hardware and firmwares (FPGA,chips) − Device drivers (may be or may be not) Don’t change clients!! 13 Goal § a prac%cal uplink without modifying client hardware or soYware. 14 Wireless Communication Primer § Packet preamble is used to measure CFO and channel coefficients − Preamble is known to the sender and the receiver a priori − The receiver compares received signals with preamble and get informa(on (maximum likelihood) Preamble Frame A Preamble sent : 1 -1 1 Preamble received : 0.9 -1.1 0.9 à CFO = 150Hz à Channel = 0.3 + 0.9j − But possible only when the preamble is delivered in the clean channel without interferences! 15 Challenges § When packets collide in random networks Frame A Decoding Failure Frame B t1 t2 t3 time − Hard to obtain channel coefficients! 16 Carrier Frequency Offset § Up-‐conversion and Down-‐conversion Up-conversion 0Hz Frequency At transmitter, data is modulated into baseband signals Down-conversion F (Carrier) The baseband signals are loaded onto carrier (up-conversion) and transmitted over the air 0 + Δf At receiver, received signals are down-converted into baseband signals again § During the down-‐conversion, clock discrepancy between the sender and receiver causes carrier frequency offset 17 Wireless Communication Primer § Carrier Frequency Offset At the transmitter, Signal : 1, 1, 1… Time 1/F 1/F At the receiver Signal : 1, 0.8, 0.4… 1 / (F + Δf) 1 / (F + Δf) 18 Challenges § When packets collide in random networks Frame A Decoding Failure Frame B t1 t2 t3 time − Hard to obtain cri(cal parameters (channel coefficients, CFOs, (ming offsets, collision points) of each packet using exis(ng techniques! − Without those parameters, it is impossible to apply MIMO decoding § Problems to Solve − To obtain the parameters − Channel coefficients, collision points and CFO 19 AIMI: ACK Handling § In CSMA/CA based MAC protocols, ACK should be sent by the receiver aYer each successful packet recep8on − How should be ACK handled for mul(ple decoded packets? Frame A ACK SIFS § In MUC, we need to ACK mul8ple packets. ACK Frame A Frame B time 20 AIMI: AP-‐Side only MIMO § MIMO-‐based Interference Suppression and Cancella8on − Decodes collided packets without es8ma8ng channel coefficients § Frequency Domain Correla8on − Obtains cri8cal parameters (CPs and CFOs) from overlapped signals … AIMI AP Frame A FDC Obtain cri(cal parameters (TO & CFO) Frame B Collision IS & IC Decode and extract each packet 21 AIMI: Interference Suppression and Cancellation § Interference Suppression − Adjust the power weights of receiving antennas such that impact of interferences is minimized − Explicit informa(on of channel coefficients is not needed − Can be used to decode collided packets xi y1 Unknown Interference Receiver Frame A time xA y2 − Power weights W for y1 and y2, respec(vely − XA = W(1)·∙y1 + W(2)·∙y2 22 AIMI: Interference Suppression and Cancellation § Then, how to obtain W (assume CFO and CPs are known) − Conven(onal technique (MMSE) not applicable to collided packets − Rather, Recursive Least Squares (RLS) filter is used ü A linear adap(ve filter that generates W when interference coexists ü Requires long training sequences (30 ~), but the packet preamble is already longer y1 xi Unknown Interference Preamble Receiver Frame A time xA y2 ü RLS computes W from the training sequence 23 Fortunately, WiFi and ZigBig use…. § Preambles with good auto-correlation and cross-correlation properties and also long enough to apply successfully RLS. 24 AIMI: Interference Suppression and Cancellation § Constraint − Channel filter is built only during the preamble recep8on − A.er training, a new user cannot transmit. Frame A Frame A Frame B Only B is decodable Frame B Only A is decodable à Interference Cancella8on is performed when interference suppression is not feasible − Always at least one packet (Frame B) that can be decoded using IS Frame A Now, Decoded can decode! using IS Frame B 25 Any sequences of collided packets can be decoded! Frame A Frame A Frame B Frame C Frame B Frame C Frame D Frame E Frame D Frame E 26 AIMI: Frequency Domain Correlation § In order to perform IS and IC 1. CFO and 2. Collision points should be iden(fied à Two dimensional search problem in frequency and (me domain Frequency Domain Time Domain § But conven(onal techniques (correla(on, maximum likelihood, etc) only works when either one of CFO or CP is known § Not applicable when both are unknown 27 AIMI: Frequency Domain Correlation § U8lize frequency domain informa8on of the packet preamble 1. Convert and keep the FFT result of a packet preamble Known Preamble 2. When signals are detected, perform FFT over n samples (n : length of the preamble) 28 AIMI: Frequency Domain Correlation 3. Correlate those two while moving the star(ng posi(on Preamble Interference Frame A t2 time 4. Find correla(on peaks 29 AIMI: Frequency Domain Correlation § The FDC output gives two informa8on 1. Collision point 2. CFO CP CFO (KHz) 30 AIMI: Frequency Domain Correlation § Why is this possible??? 1. Distribu(on of frequency spectrum is not affected by CFO ü Exponen(al changes of amplitudes in (me domain ü But only linear shiks in the frequency domain CFO 0Hz 0 + Δf Frequency 2. Packet preambles have good autocorrela(on property also in frequency domain ü Contribu(ons of other packets or unknown interference are easily cancelled Autocorrelation of packet preamble in frequency domain 31 AIMI: ACK Handling § In CSMA/CA based MAC protocols, ACK should be sent by the receiver aYer each successful packet recep8on − How should be ACK handled for mul(ple decoded packets? Frame A ACK SIFS § In MUC, several ways are possible 1. One ACK for the last packet ACK Frame A Frame B time ü Easiest, but incurs unnecessary backoff of other packet senders (Frame B) 32 AIMI: ACK Handling 2. Block ACK for the packets Frame A Frame B Block ACK for A and B Frame C ü ü Block ACK is not specified in the standards (802.15.4 and 802.11) If no-‐ACK op(on is turned on, possible to implement in sokware driver 3. Downlink beamforming of ACKs (Aryafar et al., Mobicom 2010) ACK Frame A Frame B Frame C ü ACK ACK Requires modifica(ons of clients (The AP needs CSI feedback from clients) 33 AIMI: Implementation § USRP2s with 8 RFX2400 daughter boards (2.4GHz) − Each USRP is connected to a separate PC − Due to the low processing power of USRPs, received signals are stored in the local disk and processed offline § Testbed setup SNR distribution ZigBee : -5 ~ 10 dB WiFi : 2 ~ 20 dB CFO distribution ZigBee : -20 ~ 15 KHz WiFi : -50 ~ 10 KHz ZigBee WiFi 34 AIMI: Baseline Performance § A collision of two packets received with two antennas − BER, compared with a na(ve receiver − Normalized Throughput (1 – FER), compared with a SIC and a na(ve receiver ü A SIC and a na(ve receiver works only when the SNR difference between two packets is huge. 35 AIMI: Baseline Performance § Capacity of AIMI − Almost linearly increases with the number of receiving antennas 36 AIMI: Performance in testbed § Scenarios Scenario Channel Access ACK scheme 1 CSMA/CA Last packet 2 CSMA/CA Last packets within SIFS 3 CSMA/CA All packets 4 p-persistent CSMA N/A − Scenario 1~3 : Basic CSMA/CA with 3 different ACK schemes 1. Only the last packet among the decoded packets is counted as goodput 2. Decoded packets that fall within SIFS interval from the lastly received packet 3. All the decoded packets − Scenario 4 : p-‐persistent CSMA with an op(mal p ü ü In ZigBee test, p = 0.2083 was used In WiFi test, CW was fixed to 8 (p = 0.0937) 37 AIMI: Performance in testbed § Without hidden terminals − No significant difference between a na(ve receiver, SIC, ZigZag and SAM − AIMI outperforms others because it can perform interference suppression even when packets are aligned 38 AIMI: Performance in testbed § With hidden terminals − Overall performance drop for a na(ve receiver, SIC and ZigZag because of increased collisions − Some performance improvement with SAM, since packets from hidden nodes are not aligned − Performance of AIMI also increases because of increased collisions 39 Conclusion § AIMI improves network performance only changing AP − About 100% throughput improvement without any change in the clients § If client changes are allowed, − 900% improvement for ZigBee and 500% improvement for WiFi 11Mbps § Issues − Rate adapta(on ü Complicated tradeoff between data rate and SNR à future work − Applicability to OFDM systems ü Feasible, since OFDM systems occupy similar preamble structure − Complexity ü FDC : O(c) à only depends on the length of preambles ü IS and IC : O(n) à depends on the length of packets 40