Redshift: The Explosion of Massive Scale Systems by Dr. Greg Papadopoulos

ANALYST SUMMIT 2007

Dr. Greg PapadopoulosExecutive Vice President, R&DChief Technology Officer

Redshift: The Explosion ofMassive Scale Systems

These slides contain forward-looking statements regardingthe future results and performance of Sun Microsystems, Inc., including statements regarding future industry trends and growth, demand for HPC and future trends in computing. These forward-looking statements involve risks and uncertainties, and actual results could differ materially from those contained in these forward-looking statements. Factors that could cause actual results to differ materially from those contained in these forward-looking statements include: risks associated with developing, designing, manufacturing and distributing new products; lack of success in technological advancements; pricing pressures; lack of customer acceptance of new products; the possibility of errors or defects in new products; competition; adverse business conditions; failure to retain key employees; the cancellation or delay of projects; our reliance on single-source suppliers; risks associated with our ability to purchase a sufficient amount of components to meet demand; inventory risks; risks associated with our international customers and operations; delays in product development or customer acceptance and implementation of new products and technologies; our dependence on significant customers and specific industries; and our dependence on channel partners. Please also refer to Sun's periodic reports that are filed with the Securities and Exchange Commission, including Sun's annual report on Form 10-K for the fiscal year ended June 30, 2006 and its quarterly report on Form 10-Q for the fiscal quarter ended October 1, 2006. Sun assumes no obligation to, and currently does not intend to, update these forward-looking statements.

25 Yrs: 100,000 to 1M Times More MIPS/$

1880 19201900 1940 1960 1980 2000 202010-15

10-10

10-5

100

105

1010

Year

MIPS

/1000

USD

0.01 MIPS/$1K

10,000MIPS/$1K

Hermann Brunner, Max-Planck-Institut fuer Extraterrestriche Physik, Germany

Where is the Demand?

1985 1995 2005 20151990 2000 2010.1

1

10

100

1K

10K

1M

World GDP

ProcessorPerformance

$36T$17T

Sun 11 MIPS

$40K (2005$)

T100010,000 MIPS$5K (2005$)

Core per Enterprise Demand

1000x

.1

1

10

100

1K

10K

100K

1M

20101995 2000 2005

Core Enterprise Demand

Moore’s Law

“Last Mile” Bandwidth

1K

10K

100K

1M

10M

100M

1G

10G

1985 1995 2005 20151990 2000 2010

9.6kbps14.4k

28.8k

56k 384k1.55m

4m

22m50–100m

POTS

DSL/Cable

Fiber

ΣBW – Filling the Pipe

NingΣBW

In = Out

ΣBW – Filling the Pipe

NingΣBW

In = Out

# Servers * BW/Server = # Devices * BW/Device

ΣBW – Filling the Pipe

.1

1

10

100

1K

10K

100K

1M

1995 2000 2005 2010

1,0001,000

50

1M ΣBW

Moore’s Law

Core

HPC – Insatiable Demand

HPC – Insatiable Demand

1

10

100

1K

10K

100K

1M

1995 2000 2005 2010.1

HPCHPC

Moore’s Law

Core

ΣBW

*-Prise

Internet

WorkdayNetwork.com

*Prise

Moore’s Law

*-Prise

1

10

100

1K

10K

100K

1M

1995 2000 2005 2010.1

ΣBW

CRM Core

HPC

Redshift – a Move to Massive Scale

1

10

100

1K

10K

100K

1M

1995 2000 2005 2010.1

*Prise

Moore’s Law

ΣBW

Core

HPC

Diverging Concerns

1

10

100

1K

10K

100K

1M

1995 2000 2005 2010.1

*Prise

Moore’s Law

ΣBW

CoreConsolidation

CostServices

HPC

Diverging Concerns

1

10

100

1K

10K

100K

1M

1995 2000 2005 2010.1

ScaleEfficiencyProducts *Prise

HPCHPC

Moore’s Law

ΣBW

CoreConsolidation

CostServices

“The Google Question”

If all you cared about was massive scale, what would you build and how would you operate it?

An Answer

With “Brutal Efficiency”• Hardened metrics (utilization, power, security)• Predictability of service level• Idea-to-deploy time and productivity

An Answer

With “Brutal Efficiency”• Hardened metrics (utilization, power, security)• Predictability of service level• Idea-to-deploy time and productivity

Deep, At-scale Engineering of Systems

How Deep?

Efficiency and predictability at massive scale are as

mission-criticalto Redshift as

RAS has been tothe core enterprise

Don’t Confuse

Commoditization ofComputers

Commoditization ofComputing

Don’t Confuse

Commoditization ofComputers

With

System Includes...

Base HW Plant(Servers, Storage and Switches)

O/S Instances

Network Service

Core Services and Platforms

System Includes...

Base HW Plant(Servers, Storage and Switches)

O/S Instances

Network Service

Core Services and PlatformsMassivelyParallelSystem

SMPs Are Back (With a Vengeance!)

64 Threads in 1997E10K

SizeFull RackPower

(Systems at peak utilization)9620 Watts

Weight1,800 lbs.Performance~150k tpm

SMPs Are Back (With a Vengeance!)

z

64 Threads in 1997E10K

64 Threads in 2007Niagara 2

SizeFull RackPower

(Systems at peak utilization)9620 Watts

Weight1,800 lbs.Performance~150k tpm

SMPs Are Back (With a Vengeance!)

z

64 Threads in 1997E10K

64 Threads in 2007Niagara 2

SizeFull Rack 1UPower

(Systems at peak utilization)9620 Watts 410 Watts

Weight1,800 lbs. 45 lbs.Performance~150k tpm ~300k tpm

24x

40x

30x

2x

SMPs Are Back (With a Vengeance!)

z

64 Threads in 1997E10K

64 Threads in 2007Niagara 2

Neptune and Crossbow

• Solaris virtual networking stacksfor high speed networks, based on protocol, service, or container

• Does the “careful” unleashing of 10G

• A shared, virtualized, non-blocking,multi-homed 10 Gigabit EthernetPCI-e network interface

• Gets the most out of your systemand your I/O bus> Processors get faster faster than I/O buses get

faster – PCI-e is (or soon will be) your bottleneck

Crossbow

Neptune

Neptune and Crossbow

SolarisCrossbow and PEF

MultiCore Microprocessor

PCIe4x10GE

Virtualized Network Streams

1000’s Logical Connections

Neptune

Solaris Zones Processes

Java RTS + Solaris =Open Real Time Operating System

Application

Java RTS 2.0

Solaris

Hardware

Java RTS + Solaris:1000 Times More Predictable

20microseconds

10milliseconds

12milliseconds

10microseconds

5milliseconds

7milliseconds

Maximum Latency Jitter

RTS

Real-Time Application Server ResultsComparison of Conventional Java SE to Java RTS

Distribution of Request/Reply Round-trip Times

0 200,000 400,000 600,000 800,000

10,000

1,000

100

10

1

Round-trip Time (US)

Java RTS Results: ALL priority transactionscomplete in 11 milliseconds or lessStandard Java SE Results: Prioritytransactions range from 11 ms to 3.5 seconds

Numb

er of

Req

uests

Real-Time Javaon Solaris:

Inverted PendulumGreg Bollella

Distinguished Engineer,Sun Microsystems

NetBeans IDE Support

• Cross-develop onthe host

• Deploy over the network

• Execute on the target• ...from the

NetBeans IDE

Download the JavaRTS Plugin, and

Co-Design

Computer+ Storage+ Network+ Power+ Cooling+ Software= BLACKBOX

But Wait, There’s More• The next wave will include leading-edge enterprises• Critical to this wave will be the bridge between

core IT and redshifting services:

Identity and securityProcedural languages and scripting

SOA and Web 2.0New clients

The Bridge

1

10

100

1K

10K

100K

1M

1995 2000 2010.1

*PriseHPCHPC

ΣBW

IdentityRSWSNew

Clients

Enterprise Bridge

Internet

Workday

CoreEnterprise

Network.com

Putting it all Together:The Binary Distribution

Base HW Plant(Servers, Storage and Switches)

O/S Instances

Network Service

Core Services and Platforms“TheDistro”

Source vs. BinaryPlatform Developers

Platform(Open Source)

open Javaopen ...

Source vs. BinaryPlatform Developers

Platform(Open Source)

1.1 1.2

Distributionopen Javaopen ...

1.2.1

...Binary Distribution

Source vs. BinaryPlatform Developers Application Developers

Platform(Open Source)

Applications

1.1 1.2

DistributionApplication Binaries

Binary Distribution

open Javaopen ...

1.2.1

...

Big Computers100Ks – 1Ms NodesGlobally Distributed

Devices

A Likely Scenario

Live.comGCC

Big Computers100Ks – 1Ms NodesGlobally Distributed

Distribution Networks

Devices

A Likely Scenario

Live.comGCC

Big Computers100Ks – 1Ms NodesGlobally Distributed

Distribution Networks

Startups

Devices

A Likely Scenario

Live.comGCC

Thank You.

ANALYST SUMMIT 2007

The Network is the Computer.TM

Dr. Greg PapadopoulosExecutive Vice President, R&DChief Technology Officer