a(framework(for(deploying(( signal(processing(applicaons ... · dsp1 dsp2 dsp3 vcp0(viterbi...

Atomix: A Framework for Deploying

Signal Processing Applica:ons on Wireless Infrastructure

Manu Bansal, Aaron Schulman, Sachin KaA

Stanford University

NSDI ‘15

SoIware-‐defined base-‐sta:ons

int main() { //PHY & MAC … }

Wireless infrastructure is programmable

TI 6670 mul:core DSP SoC

DSP0 DSP1

Shared SRAM

Local SRAM

Local SRAM

We could be deploying apps into infrastructure

3

RF localiza:on (SecureArray)

Video-‐op:mized wireless stack

(APEX)

BER feedback (SoIRate)

What modifica:ons are needed to deploy apps?

Primi:ves needed for deploying apps

4

Base-‐sta:on soIware needs to provide a modular interface to tap, tweak, and insert

CSI

EQ VITDEC

Tap

OFDM SLICER

(BER)

Insert

SYNC

Tweak

Apps require high throughput and low latency

5 SoIware must deliver hardware-‐like performance

Time 250us 16us

Data frame ACK frame

… 20Msps == 640Mbps

High processing throughput

Low processing latency

WiFi stack example (LTE has easier latency constraints)

GeAng hardware-‐like performance

6

DSP0 DSP1

Shared SRAM

Local SRAM

Local SRAM

Pipeline programming Memory management Inter-‐core data transfers

Customized hardware

Hand-‐op:mized soIware


7

DSP0 DSP1

Shared SRAM

Local SRAM

Local SRAM


Customized hardware



8

DSP0 DSP1

Shared SRAM

Local SRAM

Local SRAM


Modular changes can have unpredictable effects on :ming

Customized hardware


The Atom Abstrac:on

9

Atom A

tA (e.g. 200 cycles)

Atom: A unit of execu:on with fixed, known :ming

Atom A Atom B

tA + tB

Composability: Composi:on of atoms is also an atom

Base-‐sta:on soIware in Atomix… 1)  Can be built en:rely out of atoms

2)  Achieves hardware-‐like performance 3)  Enables modular modifica:ons

Base-‐sta:on soIware can be built with atoms Atomix WiFi chain can meet throughput and latency Apps are easily added to Atomix WiFi

10

WiFi signal processing chain

11

Time

Data frame Data frame

…

CSI EQ

OFDM SLICER SYNC

Data flowgraph over signal processing blocks

(BPSK or QPSK)

(BPSK) (QPSK)

Implemen:ng blocks as atoms

12

Slicer QPSK

BPSK Constella:on? BPSK: 200cy QPSK: 400cy

Input data length? 1 OFDM symbol: 200cy 2 OFDM symbols: 400cy

N? M?

1)  Split out branches, 2) Fix data lengths Make branching explicit

tBPSK48 = 200cy

QPSK48

48 cplx 48 bits

48 cplx 96 bits

BPSK48

CSI EQ

OFDM SLICER SYNC

(BPSK or QPSK)

Implemen:ng flowgraphs as atoms

13 Make branching explicit

CSI EQ

OFDM SLICER SYNC

CSI EQ

OFDM BPSK48 SYNC

CSI EQ

OFDM QPSK48 SYNC

(BPSK or QPSK)

Implemen:ng flowgraphs as atoms

Explicitly model data access cost using FIFO atoms 14

BPSK Atom

CSI EQ

OFDM BPSK48 SYNC

CSI

EQ

OFDM

BPSK SYNC

F F

F F

F F

F F F F F

BPSK F F

Parallelizing flowgraphs with atoms

15

BPSK Atom CSI

EQ

OFDM

F F

F F

F F

F SYNC F F

Core 0

Core 1

BPSK F F

Parallelizing flowgraphs with atoms

16

EQUALIZER Atom – Core 0

BPSK48 F F

BPSK Atom – Core 1

Transfer F F

CSI

EQ

OFDM

F F

F F

F F

F SYNC F F

Core 0

Core 1

Explicitly model data transfer cost as an atom

Implemen:ng decisions with atoms

17 Make branch explicit, push to the top-‐level

DISPATCHER Header atom

BPSK atom

QPSK atom

BPSK

QPSK

Atomix framework in 1-‐slide

•  Everything as an atom – Signal processing components – Hardware management components

•  Atoms for blocks, flowgraphs, states •  Simple control flow makes atoms composable •  Declara:ve language allows easy modifica:on

18 Atoms enable modularity, precise :ming, efficient pipelines


19

Fine-‐grained pipeline parallelism

20

CSI

EQ

OFDM

QAM64

SYNC

DEINTER-‐LEAVER

DEPUNC-‐TURER

DECODE-‐SCATTERER

VITERBI-‐ ISSUE

DECODE-‐GATHERER

DESCRA-‐MBLER CRC32

DSP0

DSP1

DSP2

DSP3

VCP0

VITERBI-‐DECODING

VITERBI-‐DECODING

VITERBI-‐DECODING

VITERBI-‐DECODING

VCP1 VCP2 VCP3

80-‐sample buffers (OFDM symbols)

Decoded bits … 010110

…

22

Packet Energy

CRC Toggle

Fine-‐grained pipeline parallelism

23 Atomix WiFi decodes 10MHz with resources to spare

Tight packet decode latency

24

WiFi highest-‐MCS 1000-‐byte packets, CDF of decode latency (deadline = 64us at 5MHz, 32us at 10MHz, 16us at 20MHz)

Atomix WiFi decodes 10MHz in low latency, predictable :ming

Experience with WiFi in Atomix

25

Low-‐level code (C)

Signal processing func:ons (C)

Parallelized Atoms (Ax)

Schedule, Resource Assignment

Atomix compiler

Atomix run:me libraries

Na:ve app binary

Na:ve compiler

3,000 Loc (with Atomix)

30,000 LoC (w/o Atomix)


26

Loca:on-‐signature app

No change in WiFi packet decode latency 27

CSI EQ

OFDM

SLICER

PH

SYNC

Rxx EIG SPT

4 new signal processing blocks 30 lines of code to add app

Predictable change in SYNC atom

2 new signal processing blocks 20 lines of code to add app

Predictable change in SYNC, DATA

CSI EQ

EVM

OFDM

SLICER

CIR

SYNC

ED

CNR

Related work •  Modular frameworks for GPPs and FPGAs –  SORA: works on GPPs, no clear mapping to DSPs – AirBlue: Targets FPGAs, different challenges than DSPs –  Ziria: complementary to Atomix

•  Embedded real-‐:me opera:ng systems (Neutrino, VxWorks, TI SYS/BIOS) –  Typically for low sample rate apps (e.g., an:-‐locking brakes)

– Misfits for expressing modular signal-‐processing apps – No abstrac:ons for blocks, flowgraphs, state-‐machine

28

Conclusion Atomix: a new programming framework •  Everything is an atom •  Hardware-‐like performance •  Modularity to tap, tweak, insert Future work: •  Automated resource scheduling •  Sta:c program checking •  Extending to L2-‐L7 NFV packet processing

29

a(framework(for(deploying(( signal(processing(applicaons ... · dsp1 dsp2 dsp3 vcp0(viterbi...

Documents