Real- Time Wireless Sensor- Actuator Networks

(1)

Real-‐Time Wireless Sensor-‐

Actuator Networks

Abusayeed Saifullah

Department of Computer Science

(2)

Roadmap

Ø

 

Real-‐time transmission scheduling theory for WSAN

q  Dynamic priority

q  Fixed priority

-‐  End-‐to-‐end delay analysis

-‐  Priority assignment

Ø

 

Scheduling-‐control co-‐design for WSAN

Ø

 

CapNet: real-‐time WSAN for data center power capping

(3)

42% 5% 53% Power Other Server

Wireless Data Center Power Management

Power infrastructure in an

enterprise data center costs

hundreds of millions USD.

[C our tes y: J H ami lton , Ama zon]

Power capping

: regulate the power

consumption of a cluster of servers.

(4)

42% 5% 53% Power Other Server

Wireless Data Center Power Management

Power infrastructure in an

enterprise data center costs

hundreds of millions USD.

Management using low-‐cost wireless

q  Reduced cost: for 100,000 servers, only

5.4%-‐8% of the wire solution cost

q  Simple, Ulexible, easily manageable

[C our tes y: J H ami lton , Ama zon]

Power capping

: regulate the power

consumption of a cluster of servers.

www.s of tlay er .c om

(5)

Power capping: brings the aggregate power consumption back to cap Time Power Nameplate rating Circuit allocation Actual power consumption

Power Capping

Ø

 

Cap:

circuit breaker’s capacity for a cluster of servers

Oversubscription:

put more

servers on a circuit

Servers rarely operate at peak

Aggregate power rarely exceeds

the cap

Ø

 

Using nameplate power rating à less servers per cluster

(6)

Power capping

must be done within trip time

(

deadline

)

Power Capping Real-‐Time Requirement

Ø

 

Trip time

: allowed time

duration above the cap

Ø

 

If oversubscription

duration > trip time à

circuit breaker trips

q 

Server shutdowns

q 

Power outage

Not Tripped Tripped Long-delay Conventional Tripping Short Circuit

(7)

Designing Wireless Capping Protocol

Ø

 

A

naive protocol

will repeatedly

q  Collect power consumption data from each server

q  Determine aggregate consumption; do capping if it exceeds cap

(8)

Time

Power

Designing Wireless Capping Protocol

Ø

 

A

naive protocol

will repeatedly

q  Collect power consumption data from each server

q  Determine aggregate consumption; do capping if it exceeds cap

Ø

 

We design a distributed

event-‐driven protocol

q  Global cap to local cap à local detectionà servers generate alarms

q  Aggregation is done only upon detection of a potential event

Interferes with wireless in other clusters/applications!

Power capping is a rare event!

(9)

Event detection phase max. num. of iterations done? yes no Aggregation phase Control phase start

CapNet Design and Implementation

Ø

 

CapNet

: Wireless Sensor

Net

work for Power

Cap

ping

q  Employs conUlict-‐free, event driven protocol

Ø

 

Operates in 3 phases

q  Local detection à alarm

q  Aggregation upon k alarms: increase k to suppress false alarms

q  Triggers capping, if the alarms are true

CPU frequency modulation

(10)

Event detection phase max. num. of iterations done? yes no Aggregation phase Control phase start

CapNet Design and Implementation

Ø

 

CapNet

: Wireless Sensor

Net

work for Power

Cap

ping

q  Employs conUlict-‐free, event driven protocol

Ø

 

Operates in 3 phases

q  Local detection à alarm

q  Aggregation upon k alarms: increase k to suppress false alarms

q  Triggers capping, if the alarm is true

CPU frequency

modulation

_Implemented

in TinyOS

on TelosB

platform.

(11)

Power Consumption in a Data Center

180 servers (2880 readings per server)

Rack16 Cluster 2 | Rack13 Cluster 1 | Rack8 Cluster 1 20 40 60 80 100 120 140 160

20 40 60 80 100 120 140 160 −0.2 0 0.2 0.4 0.6 0.8 1 Rack8 Cluster 1 Rack13 Cluster 1 Rack16 Cluster 2

Ø

 

Strong synchrony among the servers in the same cluster

q  Cluster exceeds cap à many servers exceed local capà fast alarm

q  CapNet beneUits from correlation à a practical approach!

180×180 matrix

[i,j]: correlation between i-‐th and j-‐th server

Microso' Data Center clusters running Web-‐ Search, Email, Map-‐

(12)

Experiment

q 

Deployed in Microsoft

data center experimental

facility

q 

TelosB to Server through

(13)

−1 −0.5 0 0.5 1 x 106 0 0.2 0.4 0.6 0.8 1

Slack considering lower trip times (ms)

CDF

4 iterations 8 iterations 16 iterations

Real-‐Time Performance (480 servers)

Not Tripped Tripped Long-delay Conventional Tripping Short Circuit

94% transmission suppression compared to the naive protocol!

Capping latency= Network latency

+ OS latency

+ Hardware latency

Slack=

Trip time-‐capping latency

110-‐350ms

{

CapNet meets the real-‐time requirements in capping.

(14)

Roadmap

Ø

 

Real-‐time transmission scheduling theory for WSAN

q  Dynamic priority

q  Fixed priority

-‐  End-‐to-‐end delay analysis

-‐  Priority assignment

Ø

 

Scheduling-‐control co-‐design for WSAN

Ø

 

CapNet: real-‐time WSAN for data center power capping

(15)

WSAN over TV White Space

Ø

 

WSAN scalability challenge

q  Smart city à hundreds of thousands of nodes

q 

Clinical monitoring in large medical systems

Ø

 

Approach: WSAN over white spaces

(16)

Sensor Networking over White Space

Advantage

q  Long transmission range

q  Wall penetration

Challenge

q  Energy, real-‐time requirements

q  Exploit bandwidth and data rate

White Spaces:

unused UHF/VHF band between 50-‐698 MHz

[courtesy: Spectrum Bridge]

Spectrum inside a data center

(17)

Sensor Networking over White Space

Advantage

Challenge

White Spaces:

unused UHF/VHF band between 50-‐698 MHz

More than 60% spectra are white spaces

< -‐85dBm

(18)

0 10 20 30 40 50 0 500 1000 1500 2000 2500 3000 Channel Availability (counties) Fixed (4 W), Portable/mobile (100 mW) Portable/mobile (40 mW)

Real- Time Wireless Sensor- Actuator Networks