DISTRIBUTED SYSTEMS -

Jan M. RabaeyDonald O. Pederson Distinguished Prof.

Director FCRP MultiScale Systems Center (MuSyC)

Scientific Co-Director Berkeley Wireless Research Center

University of California at Berkeley

INTEL, FEBRUARY 23 2011

The Next Grand Challenge in Embedded System Design

Infrastructural

core

The Swarm and the Cloud

TRILLIONS OF

CONNECTED DEVICES

[J. Rabaey, ASPDAC’08]

THE

CLOUD

THE SWARM

The New Moore’s LawStill … Improve functionality per unit cost to create

whole new application areas,

But in a brand new setting

1970

Mainframes

1980

PCs

1990

Internet

2000

Wireless &

Personal Devices

2010-

Cloud

Computing

Immersive User

Experiences

Ubiquitous

Sensing

The Swarm Perspective

It’s A Connected WorldTime to Abandon the “Component”-Oriented Vision

Moore’s Law Revisited:

Scaling is in number of connected devices,

no longer in number of transistors/chip

[MuSyC 2009]

The functionality is in the swarm!

Resources can be dynamically

provided based on availability

One Vision: CyberPhysical

SystemsLinking the Cyber and Physical Words

[H. Gill, NSF 2008]

Another One: BioCyber (?) SystemsLinking the Cyber and Biological Worlds

Examples: Brain-machine interfaces and body-area networks

The Cloud and the Swarm

Distributed Sense and Control Challenges

Complexity

Modeling/

Abstractions

System

Metrics

(ENERGY)

Run-time

Management /

Diagnostics

Verification

Security/Tru

st

Robustness/

Reliability

Failure to Address in

Fundamental and

Cohesive Way will

Slow Down or Prohibit

Adoption

It’s All About Energy

Energy among most compelling

concern of distributed IT platform and

its applications.

Mobiles

Smart grid

Avionics

Human-centric

systems

OUR VISION: Distributed

Sense and Control Systems

to Dynamically Enforce

Energy-Proportionality

Business as Usual Will Not Do

The mantra’s of two decades of low-power design:

slow, simple, many, dedicated, adaptive

While some opportunities are left, concepts now commonly exploited

The end of voltage and

energy scaling !?

Unless novel devices are adapted soon …

0.001

0.01

0.1

1

0 0.2 0.4 0.6 0.8 1 1.2

Vdd (V)

En

erg

y (

no

rm.)

Total

Switching

Leakage

0 0.2 0.4 0.6 0.8 1 1.2

VDD (V)

0.001

0.01

0.1

1

Energ

y (n

orm

.)

0.3V

12x

In Need of Novel

Architectural Ideas

The Golden Opportunity

Energy-efficiency of most systems decreases under reduced loads

Energy-Proportional Computing

Thro

ughput Actual

Ideal

Power

Courtesy:

L. Barroso, Google

Computation and Energy

Throughput

Power

Actual

Ideal

DOING NOTHING (or

LITTLE) WELL

Energy efficiency of most systems degrades

under reduced load conditions

How we design systems

How nature designs systems

[* Term coined by L. Barroso, Google]

A Generic Concept

Throughput

Po

we

r

Actual

Ideal

DOING NOTHING (or

LITTLE) WELL

Conceive and Enable Systems that are Energy-Proportional over Large Throughput Range. Applies to all aspects of the IT Platform!

Not the case in today’s systems (computing, storage, communication)

The Big Picture

Hugely Scalable Platforms“Providing computation/computation at the optimal

energy”

Attention-Optimized

Computing/Communication“Matching computation to desired utility”

Utilit

y

Maxim

ization

A Closed Loop System

The Cloud/Swarm

Challenge

Trade off computation and communication in light of limited energy, communication and,

computational resources so that desired utility is reached under highly variable conditions and loads

Requires scalable distributed optimization strategy

The “Playground”

A continuously changing

alignment

(environment, density, activity)

The Swarm/Cloud Operating System -Dynamically trading off resources

The Swarm/Cloud Services and Applications

“What matters in the end is the utility

delivered to the user”

Utility Maximization

Distributed Resources

Communication

(Spectrum)Computation Sensing

ActuationStorage Energy

ADDRESSING THE

CHALLENGES

17

Focus Center Research Program

Features

Multi-university teams

Focus on topics where evolutionary R&D is insufficient

Emphasis on discovery; long-range time horizon

Large-scale effort (~ $7M per center annually)

Equal cost sharing between industry & government

Access to relevantly trained graduate students

“The (SRC) focus center program is designed to create a nationwide, multi-university network of research centers that will keep the United States and U.S. semiconductor firms at the front of the global microelectronics revolution.”

Craig R. BarrettRetired Chairman of the Board, Intel

Former Chair, Semiconductor Technology CouncilRecent Chair, FCRP Governing Council

MUSYC IN A NUTSHELL

Grand Goal:

Grand Challenge:

Create comprehensive and systematic solution the

distributed multi-scale system design challenge.

“Energy-smart” distributed systems, that

Are deeply aware of balance between energy availability and

demand

Adjust behavior through dynamic and adaptive optimization at

all scales of design hierarchy.

Common Core:

20 Faculty Distributed over 10 US Universities

SCS Theme

Distributed sense and control systems.

Target: Airborne Platforms (Avionics)

LSS Theme

Large-scale “energy-intensive”

systems

Target: Data centers

SSS Theme

Small-scale “energy-frugal” systems

Target: Human-centered networks for

augmented sensing (e.g. BMI)

Exploring the multi-scale space:

THE MUSYC TEAM

SCS

LSS

SSS

Including experts in petascale computing, networking, control, signal

processing, information theory, avionics and neuro-engineering

THEME 1: DISTRIBUTED SENSE AND CONTROL SYSTEM METHODOLOGY (A. SANGIOVANNI-VINCENTELLI)Address challenges in complex distributed control systems by employing structuredand formal design methodologies that seamlessly and coherently combine variousdimensions of multi-scale design space, and that provide appropriate abstractions tomanage inherent complexity.

Case study: Avionics

Complexity

SCS DRIVERS AND METRICS

Today Power sources/sinks

Electric distribution

Control system

• # power sources ~1• # loads ~100• peak power ~ 400kW

• # power sources ~ 10• # loads ~1000• peak power ~ 4MW

Tomorrow

Large Airborne Platforms

In Line with DARPA META Program

Reduction of development time of complex, distributed control systems by 2X through increased use of formal methods for specification, design and verification. Reduction of the number of faults that require the system to be taken out of service for inspection or repair by 2X, through the increase used of onboard models and dynamic reconfiguration to provide enhanced fault tolerance.

SCS HIGHLIGHT: FORMULATED DESIGN FLOW FOR DISTRIBUTED AVIONICS SYSTEMS

Platform-based design enables architecture exploration (tradeoff weight, stability, …)

Power SystemArchitecture

Control System Architecture

Hardware, Software, Communications

Redesign

Incremental conservative design• Steady state worst case power draw• 2x overdesign results in weight penalty

Dynamics problems identified in verification

Communications latency impacts stability

Dynamics, control, communication latency addressed in all layers

Current State of the Art

Robust design for distributed control system

Ptolemy, Metro tools enable robust design of complex dynamical systems

Our Approach(STRONG impact on META I and II BAA)

Collaboration with UTC (HS), IBM and Raytheon

Contributors: E. Lee, R. Murray and ASV

Realistic Test Benches under development

THEME 2: LARGE-SCALE SYSTEMS (T. SIMUNIC-ROSING)

Realize distributed closed-loop power-management strategies that result in “energy-intensive” large-scale systems to be orders of magnitude more energy-efficient, whileensuring that mission-critical goals are met. To be accomplished by employing holisticmulti-scale solution including all components of the system at multiple hierarchy levels.

Target: Data centers

“Doing nothing well”

LSS DRIVERS AND METRICS

SOLUTION: Distributed and hierarchical management that ensures that energy is only consumed if, when and where needed.

Enable “energy-proportional” computing, and to “do nothing well” in Datacenters and Cloud Computing

METRIC: Datacenter Energy Efficiency

Barroso & Hölzle, 2009

LSS HIGHLIGHT: ENERGY-AWARE LOAD

SCHEDULING

BWorkload Model/

Predictor

Energy AwareWorkloadScheduler

ClusterManager

Building/Facility ManagerTasks

SLAs

Energy Supply Information

Energy Consumption

Application Resource Footprint

Contributors: Katz, Snavely, Rosing, NSF GreenLight

Cooling-aware management

THEME 3: SMALL-SCALE SYSTEMS (D. JONES)

Explore absolute bounds of energy-efficiency and miniaturization in “energy-frugal”human-centric distributed IT systems, through distributed management strategy thatdynamically and adaptively selects correct operational point corresponding to varyingapplication needs in terms of accuracy or resolution.

Target: Augmented sensing in humans (BMI)

SSS DRIVERS AND METRICSKEY METRIC: UTILITY/ENERGY

Utility Maximization• Define system performance in terms of user/application relevant utility• Dynamically optimize algorithms and platforms to maximize utility

Explore, analyze, and implement advanced closed-loop learning systems in brain-machine interfaces

In collaboration with UCB Neuroscience and UCSF Neurosurgery

ScalableSignal

acquisition

Utility-OptimizingScalable Systems

Management

HugelyScalable

Processor

AttentionalAlgorithms

ScalableRadio

Frequency Tx and Rx

RF EnergyHarvesting

EfficientIntegrated

MicroscopicAntenna

Voltage ScalablePower Source

3D Integrated Packaging

HUGELY SCALABLE SSS PLATFORM

3 10 700.00

0.05

0.10

0.15

0.20

0.25

0.30

0.35

Vm

in [

V]

Stage depth [fo4]

20f

25f

30f

35f

40f

45f

Eto

tal

[J]

3

3

SSS HIGHLIGHT: HUGELY SCALABLE BMI

PLATFORMSContributors: Rabaey, Blaauw, Franzon

3D Inductors promising higher L, Q

Energy-neutral wireless link delivers energy-proportionality over broad performance range from scavenged power

3

1 mm

65 nm CMOS, in fab

[Franzon]

[Rabaey]

Low-Jitter Timers for Power Control

[Blaauw]

1.4 μJ/hour

IBM 130 nm CMOS

In Summary …

The Laws of the Cloud and the

Swarm In a connected world, functionality arises

from connections of devices.

Largest efficiency gain obtained by

balancing available resources:

computation, communication and energy.

The dynamic nature of the environment,

the needs and the resources dictate

adaptive solutions.

No one wins by being selfish.

Cooperation and collaboration are a

must.

MuSyC as a Collaborative Answer to the

Swarm and Cloud Challenges