Spa Sparsi rsity-pr propor porti tional Ba Backhaul l an and - - PowerPoint PPT Presentation

Spa Sparsi rsity-pr propor porti tional Ba Backhaul l an and Compute e fo for SDRs

Moein Khazraee, Ye

Yeswanth Gu Guddeti, Sam Crow,

Alex C. Snoeren, Kirill Levchenko, Dinesh Bharadia, Aaron Schulman

Software Defined Radios have a lot of potential

SDRs s are unive versa sal and and flexi xible

They can decode any protocol on any band: (e.g., WiFi, Bluetooth, Zigbee, FM, LTE)

SDRs s can enable exc xciting new new ap applicat cations

• ns
• Universal IoT Gateway
• Radios in the cloud

Universal Capture Flexible Compute 2.4 GHz ISM band 900 MHz band VHF Aviation Band Bluetooth/ZigBee Paging AM/FM

De Decouple pled d radio frontend from si signal-processi ssing backe kend

What makes SDRs so powerful?

Radio frontend Signal processor

ADC Filter Process

DATA

Downsample

Unfortunately, SDRs consume a lot of resources

100 MHz z (USRP 2) 800 M 800 Mbps 25 M 25 MHz Serve ver-class ss processo ssor

SD SDR backh khaul and and comput compute is s inefficient,

especially for low-bitrate frequency hopping protocols (e.g., 1 Mbps BLE)

Radio frontend Signal processor

ADC Filter Process

DATA

Downsample

Occupied Bandwidth (MHz)

512

Mbps

4

Gbps

Opportunity: the spectrum is sparsely occupied

None

SO YOU’RE TELLING ME BACKHAUL AND COMPUTE ARE THE LIMITING FACTORS, BUT THEY’RE MOSTLY WASTED ON NOISE ?!

SDRs should be sparsity-proportional

Threshold

SDRs should be sparsity-proportional

Only backhaul and compute active signals

SparSDR: Sparsity Proportional Backhaul and Compute

Radio frontend Signal processor

ADC Filter Process

DATA

Downsample Requires only residential-class backhaul

• 1. Frontend

Fits in existing FPGAs

• 2. Backend

Runs on embedded processors Sparsity-prop. Sparsity-prop.

A Raspberry Pi can handle 100 MHz bandwidth! (captured with a SparSDR-enabled USRP N210)

How can we do sparsity-proportional downsampling?

FFT 1 FFT 2

• 1. Divide the signal

into discrete intervals

• 2. Perform a Discrete

Fourier Transform (FFT)

• 3. Throw away

unused freq. bins

• 4. Backhaul what’s left

Frequency Frequency

Signal captured by SDR frontend

FFT alone is not enough

FFT spreads signals, erasing gains from sparsity

Improving backhaul efficiency with windowing

Signal captured by SDR frontend Windowing focuses energy on active bins

Improving backhaul efficiency with windowing

Cost of windowing 2x overhead

That’s the frontend. Now, the backend.

Naive ve method: Full length Inverse Fourier Transform

𝑢

Reconstructing signals on the backend

Full length IFFT 𝑔

$

Frequency Domain Time Domain 𝑢 Filtering

Full IFFT compute is not sparsity proportional

Ou Our m meth thod: : Partial IFFT

Needs additional phase correction

Partial IFFT

“Let m me s show y you t that w we c can m make existing S g SDRs s sparsity p proportional”

Frontend Implementation in FPGA

Configurable FPGA Sparsity-prop. Downsampling Packet Framer 1Gbit Ethernet 100 MHz ADC 100 MHz SDR SDR BRAM BRAM DSP DSP AD Pluto [$200] 116 (97%) 24 (30%) USRP N210 [$2K] 85 (67%) 95 (75%) USRP X310 [$10K] 772 (49%) 1417 (92%)

Opportunity: Spare resources on SDRs

SparSDR Frontend

What did we manage to fit into SDR?

US USRP N2 N210

• Max FFT length of 2048, or resolution of 48.8 KHz

AD AD Pluto

• Max FFT length of 1024, or resolution of 60 KHz

SDR SDR DSP DSP AD Pluto 24 (30%) USRP N210 95 (75%) USRP X310 1417 (92%)

Increasing FFT length improves efficiency

% of max backhaul

Challenge: There is an unpredictable data rate

Solution: FIFOs

Data rate SDR SDR BRAM BRAM AD Pluto 116 (97%) USRP N210 85 (67%) USRP X310 772 (49%)

Challenge: sample rate = FPGA clock rate

New input sample at each clock cycle

100 MHz Clock Configurable FPGA Sparsity-prop. Downsampling Packet Framer 1Gbit Ethernet 100 MHz ADC 100 MHz 100 MHz Clock 100 MHz ADC

Solution: Use pipelined implementations of FFT and multiply

I need a drink!

SparSDR Frontend

Case study: Universal IoT Gateway

• USRP N210 and Rpi 3+
• SparSDR reconstructs and decodes

Bluetooth captures in real time on Rpi 3+

Benefit of SparSDR for IoT

Is SparSDR compute sparsity-proportional?

SparSDR’s backhaul and compute scale linearly with the rate of received BLE advertising packets.

Case study:

.

Future Work

• Porting to more platforms
• Make cloud SDRs a reality
• Low cost sensors for wide deployment:

AD Pluto + Rpi ~ $200

• Make IoT Gateway more comprehensive
• Multi-protocol decoding
• Transmit side

Go get SparSDR for your N210 and AD Pluto!

It’s open source and integrated into GNURadio: https://github.com/ucsdsysnet/sparsdr