Lowcost 802 11 Wireless Infrastructure Networks Stefan Savage

Low-cost 802. 11 Wireless Infrastructure Networks Stefan Savage and John Bellardo Department of Computer Science and Engineering University of California, San Diego

Motivation • Large-scale 802. 11 deployments are expensive – Capital expenditures typically < 35% (and hardware is on commodity price curve) – Operational expenditures • • Site-survey Test and tuning Network wiring and provisioning Ongoing management (software update, rebalance, etc) • Our goal: make it cheap and trivial to provide building or campus-wide 802. 11 APs (Op. Ex -> 0) University California, San Diego – Department of Computer Science UCSD CSE

Assumptions • Radio hardware is cheap – Multiple independent radios in a package is reasonable • Antenna technology is not – Omni antennas (low gain/directional separation) • Intra/Internet access usage model – Not point to point • Largely homogenous administrative domain – Not dealing with apartment building problem (initially) • Indoor focus – 3 D, dense deployment, complex RF domain, significant spatial and temporal load shifts • Must not require 802. 11 client modifications – Ok as optimization University California, San Diego – Department of Computer Science UCSD CSE

Aside: Why use 802. 11? • Bad experience with simulation – Our wireless immigration project (USENIX Sec ’ 03) – Send CTS with large duration to freeze channel (devestating in simulation, then we built it) – Have tried three wireless simulators (including $$$) – can’t find any that predict our measurements • Multi-path, fading, variable noise, people (i. e. moving bags of water) • Variable xmit pwr, receive sensitivity, power spectrum, MAC behavior on client NICs • We want experience with real traffic driven by real users, hence we need to build real systems • Not equipped/funded to build a lot of radios – Although we do have some (Cal. Radio – at end of talk) University California, San Diego – Department of Computer Science UCSD CSE

Two elements of our work • Radio Tomography and Frequency Management (RTFM) – Measurement-based inference of RF domain capacity and contention – Auto AP configuration to maximize system goodput • Frequency, transmit power, CCA, coding – Goal: no site survey, no tuning, no manual configuration • Less. Wire – Simplified multi-hop routing (3 hop max) – Best-exit routing wrt RF domain impact – Goal: opportunistic use of wiring, expand coverage/density University California, San Diego – Department of Computer Science UCSD CSE

Radio Tomography • Key questions – If I send pkt x at rate r with power t on channel z, what is the distribution of delivery delay times? – Why? • Background interference • Co-channel interference • Client<->AP propagation (fading, multipath, etc) University California, San Diego – Department of Computer Science UCSD CSE

Radio Tomography: first try • Naïve approach – Model nodes as point transmitters with set xmit range r and channel z – If two sphere’s overlap, delay is proportional to the sum over load – Re-color, re-size to minimize delay • AP • • Why this doesn’t even vaguely work – – Non-uniform propagation Channel not exclusive Coding matters Channel conditions and clients change University California, San Diego – Department of Computer Science UCSD CSE • AP •

Radio Tomography: 2 nd & 3 rd attempts • Next idea: – Observe visible MACs and share among APs – If two nodes share the same node then assume they interfere • Problems: – Incredibly conservative (ignores attenuation and RF capture) • Next idea: – Measure RSSI and infer impact on xput – Fine grained “ground-truth” measurements (sample over every 3 x 3 feet by hand) • Problems – RSSI is very noisy and highly variable • Hard to infer “ground truth” from few samples – Very poor predictor of pkt delivery • Happy to learn about any non-brittle models here that work University California, San Diego – Department of Computer Science UCSD CSE

Radio Tomography: current approach • Synchronized Co-Channel Interference Inference – Idea: create interference and see impact (analogies to slow start) – APs send short burst on channel x at time t and rate r – Other APs measure change in re-transmission probability and backto-back xmit timing at same time (CCA indication) – Infer same from client based on retry bit in header & CRC failures • Findings – Rate sensitivity • Both for data (makes sense) and interferer (unsure) • Discontinuities (fastest rates -> practically slower) – Strong bimodality – Good at characterizing interference • ~85% for sub-second samples • Gotchas: low S/N University California, San Diego – Department of Computer Science UCSD CSE

RF Management • RF Parameter optimization (work in progress) – Minimize xmit power to maximally split offered load across APs – Color frequency and set CCA to minimize interference effects • Research questions – NP-hard, Heuristic challenge – ordering of power/frequency opt – How often to re-optimize? • Don’t want to react to short workload dynamics (ftp transfer) or RF dynamics (jiffy pop time-scale) • Client delay on reassociation – Some NICs very bad – Our APs support fast handoff (Sync. Scan, UCSD-TR) but requires client mods to take advantage • Want to react quickly to AP failure – Centralized vs distributed control? – Impact if some nodes don’t play? (e. g. static frequency inholding) University California, San Diego – Department of Computer Science UCSD CSE

Less. Wire • Idea: use additional radios to provide multi-hop backhaul • Research challenges – Point-to-multipoint route optimization over RF domain (not ad hoc routing) – Interaction with RF management • Backhaul-only channels vs joint assignment • Optimize freq/power assignment over “opportunity cost” of a route – Simplicities from being short hop (2 -3 hops max) network (very low state) University California, San Diego – Department of Computer Science UCSD CSE

Where we are • RTFM prototype limping along at UCSD – Interference inference • Background channel quality • Co-channel impact on predicted delay on given frequency • Extrapolate rate impact based on empirical curves – Greedy channel assignment based on static threshold – Lots of work left… (CCA validation, TX power, better assignment, more features to classifier) • Less. Wire – In algorithmic stage – no results to report today – We are assuming that wired bandwidth is infinite University California, San Diego – Department of Computer Science UCSD CSE

UCSD CSE Infrastructure • 266 Mhz Soekris w/40 GB trace store – Dual-radio Atheros 5212 mini. PCIs • Driver hacks for CCA adjust, per-packet TPC&rate control • Global time synced packet scheduling – 5 Ghz deployment on two floors-12 APs (soon 2 more + indoor/outdoor-40 APs) University California, San Diego – Department of Computer Science UCSD CSE

UCSD Infrastructure: Cal. Radio I • Joint project of UCSD Cal. IT 2, ECE and CSE • Intersil baseband, 2. 4 Ghz RF, DSP-based MAC (TI ‘C 5471/ARM 7, Symbol derived IP, about 4”x 4”) • Designed to allow L 2 experimentation/innovation RF PORT Keys 8 -LEDs 8 -GPO 16 b-Latch Flash SRAM 512 K x 16 256 Kx 16 SPKR MIC Optional Decode Logic PEEL +3. 3, +2. 8, +1. 8 Power Supplies RF serial bus SPI, Keys, GPIO I 2 SMc. BSP 0 Mc. BSP 1 Mc. BSP 2 Stereo CODEC TLV 320 AIC 23 SPI I 2 S 5409 DSP (100 MIPS) JTAG Test DMA External Memory Interface ALU Peripherals Amplifier 11. 2896 MHz Progra SRAM ROM 32 K x 16 16 Kmx 16 TMS 320 VC 5409 University California, San Diego – Department of Computer Science UCSD CSE

UCSD Infrastructure: Cal. Radio II • More aggressive: physical layer innovation – Several RF modules being constructed (2. 4, 5 Ghz Wi. Fi, 2 x 2 MIMO 900 Mhz, 3 -10 Ghz UWB) – Modulation all in FPGA, Matlab/Simulink compatible LVDS Serial Interfaces University California, San Diego – Department of Computer Science UCSD CSE