Stealing From an Ongoing Flow: Protocols and Prototypes Ashu Sabharwal Rice University EPFL (2007-08) Joint work with Scott Novich & Debashish Dash Microsoft Summit 2008 • Thanks to all the participants & Microsoft • Big thanks to Ranveer for putting all this together Ashu Sabharwal Rice University 7 Blind Mice QuickTime™ and a decompressor are needed to see this picture. Ashu Sabharwal Rice University 7 Blind Mice Ashu Sabharwal Rice University 7 Blind Mice Spear Cliff Fan Pillar Rope Ashu Sabharwal Rice University 7 Blind Mice Cognitive Wireless Ashu Sabharwal Rice University Cognitive Wireless • Hype or Next Big Thing ? – Feasibility ? – Extent of Utility ? – Impact as big as we will like to believe ? • Scientific questions – Relevant problem formulations – Platforms as technical demonstrators Ashu Sabharwal Rice University Outline • Testbeds/Platforms [7 minutes] – TFA – WARP • Thought Experiment to a Demo [10 minutes] – Stealing from an ongoing flow – Formulation – Result & protocol Ashu Sabharwal Rice University At-scale: TFA-Rice Mesh Network QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. • In low-income neighbourhood of Houston, Texas • TFA Charter: To empower with technology • Deployed: 4000+ real users over 4 Km2 Ashu Sabharwal Rice University At-scale: TFA-Rice Mesh Network QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. QuickTime™ and a TIFF (Uncompressed) decompressor are needed to see this picture. WARP • Current TFA speeds peak at 0.5 Mbps/user • Goal: 4-10X gains • At-speed: Use WARP for a clean-slate network Ashu Sabharwal Rice University Wireless open-Access Research Platform • WARP – Programmable FPGA platform (Virtex IIPro, Virtex 4) – High-end MIMO (upto 4x4, 60-100 Mbps) – Frameworks for clean-slate designs Ashu Sabharwal Rice University Wireless open-Access Research Platform • Multiple Design Flows – WARP + Matlab = WARPLab (offline design) – Simulink + Sysgen = WARP_Phy + WARP_MAC (real-time) – Control & Management Plane = WARPnet (deployed networks) Ashu Sabharwal Rice University WARP Users Ashu Sabharwal Rice University WARP Users (by end of Summer’08) Industry (11) • • • • • • • • • • • Xilinx (3 sites) Nokia Beijing DRS Signal Solutions Spectrum Signal Processing Irvine Sensors ASTRI (Hong Kong) Communications Research Centre Motorola Bangalore Microsoft Research Beijing Toyota Info. Tech Ericsson Research Ashu Sabharwal Academia (15) • • • • • • • • • • • • • • • UCSD UC Irvine USC Polytechnic Rutgers University of Waterloo University of Oulu Nile University RWTH Aachen University University of Klagenfurt UC Riverside UOIT UC Santa Cruz Drexel University UIUC Rice University Applications • Urban-scale mesh network deployments (TFA-Rice) – Camp & Knightly, Infocom’08 • MIMO : Sphere detection/decoding – 3G-LTE, WiMax, 802.11n (Cavallaro’s group) • PM protocols for low-power handsets – Liu and Zhong, Mobisys’08 • Cooperative communications – Random Access Cooperative Systems (Tech Report, Asilomar’08) • Cognitive wireless (today) Ashu Sabharwal Rice University Purpose of a Testbed • Verify a concept – Sanity check & feel good – Engineering approximation error • Uncover surprises – Overhead multiplier effect observed in TFA – 50X reduction in capacity due to routing packets – Need at-scale and at-speed systems for such discoveries • Thought Experiment – Mantra is “I will build” – Forces you to start with the correct setup Ashu Sabharwal Rice University Outline • Testbeds/Platforms [7 minutes] – TFA – WARP • Thought Experiment to a Demo [10 minutes] – Stealing from an ongoing flow – Formulation – Result & protocol Ashu Sabharwal Rice University Two-Flow Network Primary Rp Rs Secondary Objective: maximize rate Rs Constraint: cannot reduce primary’s rate Ashu Sabharwal Rice University Rate Region Primary Rp Cp Rp Rs Secondary Rs Cs • Since interfering links, tradeoff between their rates • True for any choice of protocols Ashu Sabharwal Rice University Rate Region hpp Rp hps Cp Rp hsp Rs hss Rs Cs • The whole region depends on topology – Topology = {hpp , hss , hps , hsp , … } • If region is known, then rate Rs is easy to find. Ashu Sabharwal Rice University Key Issue: Lack of Knowledge Primary Rp Cp Rp Rs Secondary Rs ? Cs • Compound Network: The secondary does not know – the topology – Rp • How can it select the Rs ? Ashu Sabharwal Rice University Without Help, Secondary Cannot Send Rp Rs • Without any knowledge, max Rs = 0 • Solution = Cognition – Snoop to learn – What can one learn about this region ? Ashu Sabharwal Rice University Information Content in Snooping Primary Rp Rs Secondary Silent • Hear and decode all transmissions – Estimate primary rate, Rp – eg. by listening to ACKs • Estimates are never perfect – Overhearing over noisy wireless channels Ashu Sabharwal Rice University Information Content in Snooping Primary Rp Rp Rs Secondary Silent Rs ? • Not sufficient information to estimate the region • Reason: Passive estimation – No feedback with primary • Solution: Estimation by perturbation Ashu Sabharwal Rice University Estimation by Perturbation Rp Rs + Rs Snoop • Key requirement: Primary should be adapting its rate to network conditions (e.g. TCP) • Feedback increases compound network capacity Ashu Sabharwal Rice University Estimation by Perturbation Primary reacts here Rp Rs • • • • Inject packets at a small rate See if the primary is affected If not, increase rate till it does Then adjust Ashu Sabharwal Rice University Secondary rate Protocol Trajectory R*s Ttransmit time Tsense • • • • Slow start Adapt its rate to find optimal rate Tunable parameters, Ttransmit, Tsense, Rs Work in progress: characterize convergence rate Ashu Sabharwal Rice University Secondary rate Demo on WARP Rp R*s Rs time • Primary flow alternating between high and low data rates • Secondary (estimation by perturbation) Ashu Sabharwal Rice University Secondary rate Demo on WARP Rp R*s Rs time • Primary flow alternating between high and low data rates • Secondary (estimation by perturbation) • Loss = [R*s(t)-Rs(t)]dt Ashu Sabharwal Rice University Lesson I: Starting Point • Model as if you will build it – No network information is available – Everything has to be estimated • Directly implementable without any rework – Prototype demo using WARP – Work by Scott Novich Ashu Sabharwal Rice University Lesson II: Lack of Information • Hard to steal from dumb devices (e.g. walkie talkies) – They do not react to increased interference • Easier to steal from “smart systems” – Allows one to observe their behavior by perturbing them Ashu Sabharwal Rice University Recap • Prototyping useful at many levels – Discovering surprises (TFA Network) – Thought experiment (this talk) – Sanity check (demo later) • Distributed cognitive wireless – Stealing from dumb devices not possible – Intelligently stealing from smart devices possible Ashu Sabharwal Rice University Questions ? WARP: http://warp.rice.edu TFA: http://tfa.rice.edu CMC: http://cmc.rice.edu Ashu Sabharwal Rice University