Ryan Newton , Sivan Toledo, Lewis Girod, Hari Balakrishnan, Samuel - - PowerPoint PPT Presentation
Ryan Newton , Sivan Toledo, Lewis Girod, Hari Balakrishnan, Samuel Madden Example Application: Locating Marmots + 2 Gothic, CO deployment August 2007 Voxnet Platform 2x PXA255, 64MB RAM, 8GB Flash, 802.11B, Mica2 supervisor,
Ryan Newton,
802.11B, Mica2 supervisor, Li+ battery, Charge controller
accel, GPS, Internal temp
2
with Lewis Girod & UCLA Blumstein Lab
Animal localization
3
Pothole detection Computer Vision Pipeline leak detection EEG Seizure detection Speaker identification
4
Low power sensors weak cpu/radio
Router weak cpu, strong radio Linux microserver
Mix and Match!
5
Results Sensor source(s)
6
Sensor source(s) Results
7
Sensor source(s) Results
Compile & Load Compile & Load
8
ANSI C NesC/TinyOS JavaME
Backend CodeGen
Wishbone Sample data (for profiling)
9
( , )
Execute!
tstart tend time
iterate x in S { f(); for(i=…) { … } g(); }
f() for () {…} g()
msg1 msg2 msg3
Tasks
Same goal as Protothreads
Every sensor source is
Includes timing info Measure rates,
Separately: profile
per-node send rate audioStream = IFPROF(readFile(“foo8kHz”, readSensor()))
10
3 ms 20 Kbps 27 Kbps
11
12
NP-Hard
Introduce variables where 0=server, 1=sensor Introduce variables where 1 = cut edge Enforce resource bounds where where Minimize objective function
13
u
uv Edges
Proxy for Energy Tricky bit (see paper): Relating f and g while staying linear
14
cepstrals hamming FFT filtbank logs prefilt preemph source
0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1 source preemph hamming prefilt FFT filtBank logs cepstrals Fraction of total CPU cost Operator Mote N80 PC
15
10 100 1000 10000 100000 1e+06 s
r c e p r e e m p h h a m m i n g p r e f i l t F F T f i l t B a n k l
s c e p s t r a l s 10 20 30 40 50 Execution time of operator (microseconds) Bandwidth of cut (KBytes/Sec) Cumulative CPU Cost Bandwidth (Right-hand scale)
16
Cumulative CPU cost (red) Operators:
EEG Application (1 of 22 channels)
10 20 30 40 50 60 70 80 2 4 6 8 10 12 14 16 18 20 Number of operators in optimal node partition Input data rate as a multiple of 8 kHz TmoteSky/TinyOS NokiaN80/Java
Each line represents 2100 partioner-runs
17
Speaker Detection Application 0.001 0.01 0.1 1 10 100 1000 10000 source/1 filtbank/7 logs/8 cepstral/9 Handled input rate as multiple of 8 kHz Cutpoint / number of operators in node partition TinyOS JavaME iPhone VoxNet
18
1 2 3 4 5 source hamming FFT filtBank logs cepstral Detections per second Cutpoint 1 TMote + Basestation 20 TMote Network
How many detections can we actually get out of the network?
20 40 60 80 100 source hamming FFT filtBank logs cepstral Percent Cutpoint percent input events received percent network msgs successful goodput (product)
Compute/Bandwidth Tension (1 mote + basestation)
19
Graph partitioning for scientific codes
balanced, heuristic – e.g. Zoltan
Task scheduling, commonly list scheduling Dynamic: Map-reduce, Condor, etc. Sensor network context: Tenet and Vango
Linear pipeline of operators Manual partition Run TinyOS code on both server and sensor
20
21
Graph Preprocessing step
Merge vertices until all edge-weights are
Eliminates the majority of edges
Even without preprocessing,
8000 runs, partitioning the 1400-node EEG dataflow graph, with different CPU budget, took under 10 seconds 95% of the time. But there is a long tail… luckily ILP solvers
0.1 1 10 100 1000 Seconds Time to discover optimal Time to prove optimal
22
1 2
5 4 1
1 2
5 4 1
1 2
5 4 1
1 2
5 4 1
1 2
5 4 1
1 2
5 4 1 budget = 2 budget = 3 budget = 4
bandwidth = 8 bandwidth = 6 bandwidth = 5
23