Wireless Networks Should Spread Spectrum On Demand Ramki Gummadi (MIT) Joint work with Hari Balakrishnan
HotNets The problem: Bursty traffic Demand variability observable even at short (30 s) time scales From OSDI 2006 traces Five APs, three orthogonal channels Spatio-temporal demand variations common Next 30 seconds First 30 seconds
HotNets Today: Static spectrum allocation Partitioned into non-interfering channels Avoid CSMA hidden and exposed terminals Avoid back-offs X
HotNets Insight: Spectrum tracks demand Spectrum tracking demand achieves higher SINR than shifting demand to where spectrum is
HotNets ODS: On-Demand Spectrum Demand-based spectrum to nodes Uses spread-spectrum codes Allocates multiple codes to transmitters A single transmitter can use entire spectrum
HotNets Key challenge Avoid inter-AP coordination Different admin domains Demand-communication overhead X
HotNets Mechanism: Spread-spectrum codes Data Code Signal Received signal Copy of received signal Alices code Bobs code Concurrent
HotNets Roadmap ODS design Determine demands Allocate codes Ensure conflict-freedom Use multiple codes concurrently ODS evaluation
HotNets Determining demands An AP computes demands of its own clients Averaged over last 30 s Demand if queue length q i, bit-rate r i For uplink, a client tells its queue length to AP d i = q i r i d 2 = 1 d 1 = 3
HotNets Allocating codes Large (128) codebook c of random codes Same at each AP AP allocates transmitter codes Minimizes mean transmission time. (Fairness?) i t h c i = l c d i P i d j m c 1 = 96 c 2 = 32
HotNets Code assignment Each AP assigns codes to transmitters from the codebook randomly No coordination among APs
HotNets Code selection Each transmitter selects up to k (=11, say) codes from its allocation randomly With 2 tx, 1 code, no-conflict probability: With n transmitters, 1 code, If n tx, k codes, conflict-free code number: Optimum code number as p = 1 ¡ k c p = ( 1 ¡ k c ) n ¸ = k ( 1 ¡ k c ) n ¸ o p t = c ne n ! 1 The optimum conflict-free code number under random selection within factor e of centralized The optimum conflict-free code number under random selection within factor e of centralized
HotNets Random code selection performance High throughput at low contention Non-zero throughput even with 128 interferers Random selection policy can be both efficient and robust Random selection policy can be both efficient and robust
HotNets Finding conflict-free codes Transmitter uses feedback from receiver Assign success probability p {0,1} per code Toggle p based on receiver feedback p=0 at tx whose hashed id closest to code p = 1 p = 0 p = 1 2 id=100 id=010 code=101
HotNets Using codes concurrently Divide packet into sub-packets Use one code per sub-packet Transmit all coded sub-packets concurrently Packet header tells receiver which codes are used Codes in conflict easy to identify at receiver Packet
HotNets Recap: Avoid inter-AP coordination Two key mechanisms Random code selection Efficient and robust Feedback-based conflict detection Decentralized
HotNets Roadmap ODS design Determine demands Allocate codes Ensure conflict-freedom Use multiple codes concurrently ODS evaluation
HotNets Challenge: Data reduction USRP/GNURadio USB throughput-limited Two steps needed for data reduction De-spreading and synchronization FPGA de-spreads, followed by synchronization Transmitter design similar Q Convolution Filter I Convolution Filter Rx I/Q Modem I 2 +Q 2 Peak Detector Peak I,Q Samples (USB) PC FPGA De-spreadingSynchronization
HotNets Preliminary evaluation ODS, two bonded 2 Mbps links No ODS, two bonded 2 Mbps links ODS improves link throughput by 75%
HotNets Related work Plain CDMA Inefficient spectrum usage with bursty traffic Sub-optimal Load-aware spectrum distribution (MSR) Uses channel-widths instead of codes Inter-AP coordination (10-minute updates) CDMA X l o g 2 ( 1 + P 1 P 2 + N ) l o g 2 ( 1 + P 2 P 1 + N ) VWID TDMA R1R1 R2R2 (bits/s/Hz) A B l o g 2 ( 1 + P 1 N ) l o g 2 ( 1 + P 2 N )
HotNets Contributions Exploit bursty demands to improve spectrum usage Demand-based code allocation Challenge: Avoid inter-AP coordination Random code selection Feedback-based conflict detection Future work: Better implementation, evaluation Need high-throughput, low-latency radios