Conquering Complexit y Building Syst ems wit h Billions of Part s - - PowerPoint PPT Presentation

CRA GCC ' 02 Complexit y/ Self -* 1

Conquering Complexit y

Building Syst ems wit h Billions of Part s

Part icipant s (at t he end): Rod Brooks, Set h Copen Goldst ein, Anant J hingran, Len Kleinrock, Richard Newt on, St eve Reiss, Bob Sproull

J une 25, 2002

CRA GCC ' 02 Complexit y/ Self -* 2

Conquering Complexit y

Building Syst ems wit h Billions of Part s

Part icipant s (at t he end): Rod Brooks, Set h Copen Goldst ein, Anant J hingran, Len Kleinrock, Richard Newt on, St eve Reiss, Bob Sproull

J une 25, 2002

CRA GCC ' 02 Complexit y/ Self -* 3

What do t hese have in common? musical greet ing card?

CRA GCC ' 02 Complexit y/ Self -* 4

Mission St at ement

To ref ormulat e comput ing syst ems archit ect ures at all levels (f rom circuit s t o global-scale dist ribut ed syst ems) t hat break t hrough t he complexit y wall t o deliver robust , scalable, long-last ing, syst ems.

CRA GCC ' 02 Complexit y/ Self -* 5

Why?

• We need comput ing syst ems t hat are t he

agent of change in societ y rat her t hat enemy

• f change
• Ever y ar t if act we build or grow in t he f ut ur e

will likely have a comput ing component

• We have run int o a complexit y wall, t hat limit s

and inhibit s growt h in business and societ al syst ems Tomorrow’s comput ing syst ems cannot be built using met hods of t oday.

CRA GCC ' 02 Complexit y/ Self -* 6

Two Themes

1) Complex organized behavior out of many simple unreliable component s 2) Make complex syst ems simple t o t he

– User – Administ rat or – Designer

CRA GCC ' 02 Complexit y/ Self -* 7

Common Challenges

• Federat ion of a large number of unit s
• Unit s t hat can change over t ime
• Wide and dynamic range of lat encies

and bandwidt hs among component s (all t he way t o occasionally disconnect ed?)

• Scaling wit h ease

CRA GCC ' 02 Complexit y/ Self -* 8

Common At t ribut es

• Self -conf igurat ion

– The induct ive st ep is f ree – Emergent behavior

• Self -Adapt at ion

– Changes in environment (e.g., load, f ailures)

• Reusablit y

– Small changes in f unct ion wit hout reengineering – Met a-programming

• Mot herhood and apple pie (robust , secure, st able, …

)

CRA GCC ' 02 Complexit y/ Self -* 9

How we do it now

• Abst ract ion/ Layering

– Fixed API s bet ween layers – Fixed f unct ionalit y at each layer

• Det erminist ic int eract ion bet ween

component s

• Det erminist ic approach t o f ailure

– Explicit coding of f ailure int o syst em

• Perf ormance cent ric implement at ions
• Result : Rigid, brit t le syst ems

CRA GCC ' 02 Complexit y/ Self -* 10

Possible Approaches

• Collect ive int elligence

E.g., Swarms

• Localize change
• Evolut ionary models

– Adapt at ion – Bio-mimet ic approaches

• Modeling f or syst em level ef f ect s
• More aut onomy at every level
• Market mechanisms
• Simple Many and Self -Healing (SMASH)

CRA GCC ' 02 Complexit y/ Self -* 11

Some Applicat ions

• Build reliable comput er syst ems wit h billions of

component s

• Sensor net work t hat covers t he eart h
• Connect every person t o t he net work
• Smart mat t er - r econf igurable art if act s
• Net worked mat t er (inst rument ed eart h)
• Simulat ed Realit y: Simulat ion engine
• Underst anding biological syst ems
• I nt elligent Transport at ion Syst ems
• Crit ical t o Ubicomp & Trust wort hy

CRA GCC ' 02 Complexit y/ Self -* 12

Comput at ional Paint

• Click t o add t ext

Organized behavior f rom many simple devices

CRA GCC ' 02 Complexit y/ Self -* 13

I nt elligent Transport at ion Syst ems

• Click t o add t ext

Robust , reliable, maint ainable, scalable behavior f rom complex devices

CRA GCC ' 02 Complexit y/ Self -* 14

Success I s

• Complexit y is not t he weakest link
• Met rics

– Deployed syst ems per engineer – Maint enance cost s – Administ rat ion cost s – Syst em longevit y

• At least linear improvement wit h

increased size

CRA GCC ' 02 Complexit y/ Self -* 15

Conclusions

• Complexit y limit s our abilit y t o meet

import ant needs

• Tweaking won’t solve t he pr oblem
• We need a t op-t o-bot t om re-examinat ion of

t he way we archit ect and build comput ing syst ems

We must conquer complexit y