154
© 2012 IBM Corporation sGE06 To Infinity and Beyond: 2012 Internet Scale Workloads and Data Center Design John Sing, Executive Consultant, IBM Systems and Technology Group IBM Systems Technical Universities– Budapest, Hungary – October 15-19

To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

Embed Size (px)

DESCRIPTION

To Infinity and Beyond - 2012 Internet Scale Workloads and Internet-scale Data Center Design: This Oct 2012 presentation given at the IBM Europe Symposium in Budapest, takes an advanced look at today's massive internet scale workloads and data centers, and dissects how/what lessons we can/should/must apply to our own IT shops. We'll examine how Internet Scale is very different than a collection of co-located servers - how these data centers respond to real-time, dynamic, fluid, competitive-advantage-leapfrog Internet business environments. These Internet-scale data center's servers, storage, software use new approaches to work as a end-to-end efficient, flexible, adaptable work flow. Using Google's definitive work on "The Data Center as A Computer - Intro to Warehouse-scale Machine" as a foundation (superb open license material by Google 2009), come discover the design, deployment, and lessons that we all must learn from these giants of the Internet. Why / How do they do what they do? Where are they being built? How are they powered and cooled? What are deployment form factors, design philosophies, power/cooling/packaging principles and trends, including modular portable container data center architecture? You'll come away knowing what you should apply to your own individual IT datacenter infrastructure in 2012 and beyond. My only request when using / referencing this material, is that you give full credit to me and IBM as the original authors of this research. That having been said, please spread the good word with good business judgement - we all benefit in today's modern global world.

Citation preview

Page 1: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

sGE06

To Infinity and Beyond:2012 Internet Scale Workloads and Data Center Design

John Sing, Executive Consultant, IBM Systems and Technology Group

IBM Systems Technical Universities– Budapest, Hungary – October 15-19

Page 2: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

2

IBMTECHU.COM

IBM STG Technical Universities & Conferences web portal

Direct link: ibmtechu.com/hu

KEY FEATURES...

– Create a personal agenda using the agenda planner– View the agenda and agenda changes– Use the agenda search to find the sessions and/or – Download presentations– Submit Session and Conference Evaluations

Win prizes by submitting

evaluations online. The more evalutions

submitted, the greater chance of

winning

Page 3: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

3

Evaluations are Online! Evaluations are Online! IBMTECHU.COM

sGE06

Page 4: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

John Sing 31 years of experience with IBM in high end servers, storage, and software– 2009 - Present: IBM Executive Strategy Consultant: IT Strategy and Planning, Enterprise

Large Scale Storage, Internet Scale Workloads and Data Center Design, Big Data Analytics, HA/DR/BC

– 2002-2008: IBM IT Data Center Strategy, Large Scale Systems, Business Continuity, HA/DR/BC, IBM Storage

– 1998-2001: IBM Storage Subsystems Group - Enterprise Storage Server Marketing Manager, Planner for ESS Copy Services (FlashCopy, PPRC, XRC, Metro Mirror, Global Mirror)

– 1994-1998: IBM Hong Kong, IBM China Marketing Specialist for High-End Storage– 1989-1994: IBM USA Systems Center Specialist for High-End S/390 processors– 1982-1989: IBM USA Marketing Specialist for S/370, S/390 customers (including VSE

and VSE/ESA)

[email protected]

IBM colleagues may access my intranet webpage:– http://snjgsa.ibm.com/~singj/

You may follow my daily IT research blog– http://www.delicious.com/atsf_arizona

You may follow me on Slideshare.net:– http://www.slideshare.net/johnsing1

My LinkedIn:– http://www.linkedin.com/in/johnsing

Page 5: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Agenda

Today’s Internet Scale Data Center Landscape

– Where are they? How big? How fast growing?

– What are they being used for? Cloud impact?

– Why understand them?

What is internet data center / warehouse-scale computing?

– How is it different? Workloads? – Hardware and software? – How the same?

How best to meld with it / use it / exploit?– Lessons we can apply from Internet scale

computing to traditional IT– Resources to help you on this journey

This session is the author’s research compilation, Great thanks to Google for the seminal work on which this lecture is based.Download a copy at: http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Published in 2009

Page 6: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Today’s Internet Scale Data CenterLandscape

Paraphrased:

“The world has changed. And some things that should not be forgotten, may be lost”.

The Lord of the Rings, Galadriel

Page 7: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Today: two different types of IT

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Internet scale wkloadsTransactional IT

Page 8: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Today’s two major IT workload types

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/ Transactional IT Internet scale wkloads

Page 9: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

How to build these two different clouds

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Transactional ITInternet scale wkloads

Page 10: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

What You (Consumer) Get with These Clouds:

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Page 11: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Policy-based Clouds and Design-for-fail Clouds are purpose optimized Infrastructure Management solutions

Policy-based Clouds

• Purpose optimized for longer-lived virtual machines managed by Server Administrator

• Centralizes enterprise server virtualization administration tasks

• High degree of flexibility designed to accommodate virtualization all workloads

• Significant focus on managing availability and QoS for long-lived workloads with level of isolation

• Characteristics derived from exploiting enterprise class hardware

• Legacy applications

Design-for-fail Clouds

• Purpose optimized for shorter-term virtual machines managed via end-user or automated process

• Decentralized control, embraces eventual consistency, focus on making “good enough” decisions

• High degree of standardization

• Significant focus on ensuring availability of control plane

• Characteristics driven by software

• New applications

Page 12: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

What’s happening?

Continually rising worldwide internet bandwidth

• Cisco global IP traffic study and forecast• http://www.akamai.com/stateoftheinternet/

Has given rise to pervasive and hyper-growing web services delivery model

– (i.e. “The Cloud”)

The Cloud is provided by data centers with massive amounts of well-connected processors, storage, network

– Amortized across internet scale user population– Across multiple workloads

http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns827/VNI_Hyperconnectivity_WP.html

Page 13: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Bandwidth and the Cloud…..

This new class of large-scale internet and cloud data centers

Has data volume: – 10s / 100s petabytes

Servers: – 100,000s

Workload can’t fit– In single server / rack of servers

Workload:– Requires server clusters of 100s, 1000s,

10,000s, or more…….

Page 14: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Growth ofThe Cloudby 2014

Means very big shift in resources

And in the way that IT is managed for the enterprise

http://www.cisco.com/en/US/solutions/collateral/ns341/ns525/ns537/ns705/ns1175/Cloud_Index_White_Paper.html

Source:

Page 15: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

How Big is the World? - 1

http://wikibon.org/blog/how-big-is-the-world-of-cloud-computing-infographic/

This is significant

Cheaper7.1x5.7x7.3x

NetworkStorageAdmins

Page 16: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

How Big is the World? - 2

http://wikibon.org/blog/how-big-is-the-world-of-cloud-computing-infographic/

We’re going to talk about this

Page 17: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Warehouse Scale Computers

The name for this emerging class of data centers is: – Warehouse-Scale Computer (WSC)

Large portions of hardware and software resource must work together

Only achieved by holistic approach to their design and development

Treat the datacenter itself as one massive computer

Enclosure for this computer looks like a building or warehouse

This session is the author’s research compilation, Great thanks to Google for the seminal work on which this lecture is based.Download a copy at: http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Here in plain English is the fundamentals of the next-generation IT age

Page 18: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Tell me more. How does this compare to my existing data

centers?

What are different workloads that best fit into the two different

types?

How best to meet / meld / jointly profit ?

OK. Hmmmmmmm……….

Page 19: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Big Positioning picture

TraditionalIT

Sto

rag

e r

eq’d

: G

B,

TB

, P

B

$ /

serv

er

DataWarehouse

TraditionalIT

DataWarehouse

CurrentIT

architectures

Growth areas

Mobile, Cloud

Growth areas

Mobile, Cloud

BigData

Internetscale

BigData

Internetscale

Current IT architectures

Page 20: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Build new, different skills sets

TraditionalIT

Sto

rag

e r

eq’d

: G

B,

TB

, P

B $ /

serv

er,

sto

rag

e

DataWarehouse

BigData

Internetscale

TraditionalIT

DataWarehouse

BigData

Internetscale

Current IT architectures

Traditional IT workloads

Current ITarchitectures

Highly parallelized internet scale architecture

Integrated E2E software centric

Page 21: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Key strategies

$ /

serv

er,

sto

rag

e

TraditionalIT

DataWarehouse

BigData

Internetscale

Current ITarchitectures Traditional IT

architectures

Internet scale architectures

Continue modernize current traditional IT …

Architect new-gen

connectors, skills Architect future

expandability

Connect with– New generation

mobile-enabled workloads

View new gen as powerful

partner

Enable them to view traditional IT as powerful

enabler

Page 22: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Why Warehouse Scale Computers (WSC) might matter to you

While WSCs / Cloud Data Centers might today be considered a niche area

– Their sheer size / cost / architecture is no longer uncommon

– Among large internet companies and cloud co-lo’s

Problems solved by today’s huge Internet-scale IT design-for-fail architectures

– Have already become meaningful to a much larger constituency

Many organizations are already: – Exploiting similarly architected computers, at a much lower

cost - Hadoop is an example– Soon, we may have 2000+ cores in a single server

The experience learned building today’s Internet Scale Data centers

– Is useful in preparing your team / company to meld, interact, plan, grow, expand, exploit the future in your own best interest

– As these potentially ubiquitous next-generation machines and data centers take hold

Page 23: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

What is:

Internet Scale Data Center?

Warehouse Scale Computer?

Page 24: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Internet scale scale computing- what is different?

Traditional Data Center

Co-located machines share security, environmentals

Applications = a few binaries, running on relatively small number of machines

– 100s of inter-process relationships requiring 100 nanosecond response

Heterogenous hardware, software

Partitioned resources, managed and scheduled appropriately

Facility and computing equipment designed separately

Warehouse-scale computer (WSC)

Computing system designed to run massive internet services

Highly parallelized applications = 10s of binaries running on 1000s of machines

– 100,000s of independent tasks only requiring 100 microsecond response time (100x slower)

Homogeneous hardware, system software

Common pool of resources managed centrally

Integrated design of facility and computing machinery

This is a different thing, for a different workload type

Main difference

Credit for all these ideas is to Google 2011 June talk by Luis Andre Barroso at 2011 Federated Computing Research Conference San Jose, Calif

Page 25: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Another way to tell them apart

Traditional Data Center

If your storage system has a few petabytes of data

Warehouse-scale computer

If your storage subsystem pages you in the middle of the night

Because you only have a few petabytes of free space left

Credit for all these ideas is to Google 2011 June talk by Luis Andre Barroso at 2011 Federated Computing Research Conference San Jose, Calif

Page 26: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Let’s see some of largest Internet Scale Data Centers

Many are co-location Cloud data centers

Many are true Warehouse Scale Computers

All of them have a very specific Internet web services application profile

Page 27: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://wikibon.org/blog/wp-content/uploads/2011/10/5-top-data-centers.html

Page 28: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://wikibon.org/blog/wp-content/uploads/2011/10/5-top-data-centers.html

Page 29: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Large Data Centers in past 2 years

10. SUPERNAP, LAS VEGAS, 407,000 SF

9A and 9B. MICROSOFT QUINCY AND SAN ANTONIO DATA CENTERS, 470,000 S

Page 30: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Container Data Center Architecture 7. PHOENIX ONE, PHOENIX, ARIZ. 538,000 SF

5. MICROSOFT CHICAGO DATA CENTER, Chicago 700,000 SF 2. QTS METRO DATA CENTER, ATLANTA, 990,000 SF

Microsoft’s Chicago Container Data Center

Page 31: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

More data centers….

4. NEXT GENERATION DATA EUROPE, WALES 750,000 SF

3. NAP OF THE AMERICAS, MIAMI, 750,000 SF

1. 350 EAST CERMAK, CHICAGO, 1.1 MILLION SQUARE FEET

Consumes 100 megawatts of power, 2nd-largest power customer for Commonwealth Edison, trailing only Chicago’s O’Hare Airport.

Page 32: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

2012: Other large world data centers For Tulip Telecom, India, in Bangalore

Currently largest in AP and 3d largest in world (for now)

Nearly 1 M sq feet

Co-built with IBM

China to build 6.2 M sq feet data center by 2016

Amadeus, Erding, Germany1+ billion transactions / day.3 second response timeAccess to 95% of the worlds airline seats5000+ servers Powers over 260 websites in 110 countries for over 100 airlines10 PB of storage

Utah Data Center, US Govt, 1M sq feet

Page 33: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Now….. what about the web giants?

i.e. Apple, Facebook, Google, Amazon, etc?

That’s Big!

Great Technology Wars of 2012 – Future of the Innovation Economy - Fast Company.com

Page 34: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

AppleHere’s what powers iCloud, see Jobs at WWDC 2011 iCloud announce (YouTube)

Rendering of Apple's new North Carolina Data Center. Credit: Apple

Other Apple data centers:

Cork, IrelandMunich, GermanyNewark, CaliforniaCupertion, Calif

Apple Data Center

FAQ

Maiden, North Carolina 500K sq ftUSD $1B

This is phase 1 only

Apple Data Center Newark, California

Purposes for all these data centers:

•iCloud•Support Apple’s WW install base of devices•Futures: Move Content Delivery Network in-house?•Futures: Streaming video?

Under construction: Prineville, Oregon

Page 35: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Facebook

Facebook’s North Carolina Data Center Goes Live Lulea, Sweden - 290K sq ft (27K

sq meters) by late 2012

Facebook – Prinville, Oregon

Has spent $1B on it’s data centers

Open Compute Project

http://www.wired.com/wiredenterprise/2011/12/facebook-data-center/all/1

Page 36: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Amazon

http://www.searchenginejournal.com/fathoming-amazon-a-visualization-of-their-success-infographic/36768/

Page 37: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Amazon Web Services

As of 1Q2012, AWS stores 905 billion objects and servers 650K requests/sec

Amazon Web Services 1Q12: 450,000 servers

Amazon Perdix Modular Datacenter

Amazon EC2 Cloud, with 17K core, 240 teraflop cluster is42nd fastest supercomputer in the world

450,000servers

905 billionobjects

650Kreq/sec

http://aws.typepad.com/aws/2012/04/amazon-s3-905-billion-objects-and-650000-requestssecond.html http://gigaom.com/cloud/how-big-is-amazon-web-services-bigger-than-a-billion/

Page 38: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

To understand the modern Internet scale data center……

Let’s study the creators of this new paradigm

Google originated and continues to drivemuch of this style of technology

Page 39: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

What is Google? Google is not a search engine

Google is a real-time “Data Factory” ecosystem

– Defacto organizer of all human internet data

– Provides worldwide Patterns of Life data• Search, analytics, etc as processing• Interactive maps as visualization

– Android as ingest / output devices• Motorola Wireless acquisition $12B

– Supporting businesses and ecosystem roles:• Google+, Play, Shop, Books, Gmail, Docs• Voice recognition software

The history of search engine http://www.wordstream.com/articles/internet-search-engines-history

Page 40: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Apple– Apple bought 12 PB for iTunes, iCloud– iPod = successful because of iTunes ecosystem– iPhone = successful because of App Store ecosystem

Facebook ecosystem– Patterns of life data on over 900 million users worldwide– Storage size of their Hadoop cluster: 30 PB

Amazon Web Services ecosystem

– Building 4 new modular datacenters: Oregon + Ireland– http://www.datacenterknowledge.com/archives/2011/03/28/amazons-cloud-goes-modular-in-oregon/– http://www.slideshare.net/AmazonWebServices/best-practices-in-architecting-for-the-cloud-webinar-jinesh-varia

eBay ecosystem– 2009: Analyzes 50PB of data a day, over 8 billion URL requests per day

Bottom line: ecosystem is no longer optional, hasn’t been for some time

Internet scale data centers… are “Data Factories with ecosystem”

Page 41: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google has already gone through three major end-to-end transformations

Google has 3 ages in terms of managing data:

Batch: Indexes calculated every month (2003)– Crawled the web 1x month. Built a search index, answered queries. Largely read-

only, pretty easy to scale. This is still how most people have in their minds about how Google works

Warehouse: the datacenter as one huge computer (2005) – Things move faster. The Internet has happened - pervasive, high speed,

interactive. – Building their own datacenters, more sophisticated at every level – Iconic systems like BigTable in production– At this time Google realized they were building something qualitatively different

than anything before, something we now think of as cloud computing. Amazon's EC2 launched in 2006

Instant: make it all real-time (2010)– Google's Colossus Makes Search Real-Time By dumping MapReduce– 3 Billion Writes and 20 Billion Read Transactions daily

Reflectsthe shift to

mobile devices and computing

http://www.google.com/insidesearch/features/instant/about.html

http://highscalability.com/blog/2011/8/29/the-three-ages-of-google-batch-warehouse-instant.html

The history of search engine http://www.wordstream.com/articles/internet-search-engines-history

http://highscalability.com/blog/2010/9/11/googles-colossus-makes-search-real-time-by-dumping-mapreduce.html

Page 42: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Why Google Instant? It was part of the smartphone explosion of value across Internet….

In 2011, every 5 minutes = 250 hours of YouTube video uploads

Page 43: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

You’ve noticed Google Instant:

Page 44: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Architectural Guiding Principles

ForInternet Scale Servers in

Big Data companies

Page 45: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Let’s examine the infrastructure

Looking for lessons

Hint: what is an Internet workload?

Page 46: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Internet Scale Workload Characteristics - 1

Embarrassingly parallel Internet workload

– Immense data sets, but relatively independent records being processed• Example: billions of web pages, billions of log / cookie / click entries

– Web requests from different users essentially independent of each over• Creating natural units of data partitioning and concurrency• Lends itself well to cluster-level scheduling / load-balancing

– Independence = peak server performance not important– What’s important is aggregate throughput of 100,000s of servers

i.e. Very low inter-process

communication

Workload Churn

– Well-defined, stable high level API’s (i.e. simple URLs)– Software release cycles on the order of every couple of weeks

• Means Google’s entire core of search services rewritten in 2 years– Great for rapid innovation

• Expect significant software re-writes to fix problems ongoing basis– New products hyper-frequently emerge

• Often with workload-altering characteristics, example = YouTube

Page 47: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Internet Scale Workload Characteristics - 2

Platform Homogeneity– Single company owns, has technical capability, runs entire platform

end-to-end including an ecosystem– Most Web applications more homogeneous than traditional IT– With immense number of independent worldwide users

1% - 2% of all Internet requests

fail*

Users can’t tell difference between Internet down and

your system down

Hence 99% good enough

*The Data Center as a Computer: Introduction to Warehouse Scale Computing, p.81 Barroso, Holzle

http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Fault-free operation via application middleware– Some type of failure every few hours, including software bugs– All hidden from users by fault-tolerant middleware– Means hardware, software doesn’t have to be perfect

Immense scale: – Workload can’t be held within 1 server, or within max size tightly-clustered

memory-shared SMP– Requires clusters of 1000s, 10000s of servers with corresponding PBs

storage, network, power, cooling, software– Scale of compute power also makes possible apps such as Google Maps,

Google Translate, Amazon Web Services EC2, Facebook, etc.

Page 48: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Server, storage architecture at internet scale

Internet scale server, storage architecture fundamental assumptions:

– Distributed aggregation of data

– Storage functionality is in software on the server

– Time to Market is everything• Breakage = “OK” if I can insulate that from user

– Affordability is everything– Use open source software where-ever possible

– Expect that something somewhere in infrastructure will always be broken

– Infrastructure is designed top-to-bottom to address this

All other criteria are driven off of these

Storage criteria:

Cost

Extreme:

- Scale- Parallelism- Performance- Real time-Time to Market

Page 49: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

SERVER HARDWARE

RHEL 2.6.X PAE

RACK

INTERIOR NETWORK IPv6

GFS / GFS II

BigTable MapreduceBigTable

Chubby Lock

GOOGLE APP ENGINE

Python, Java, C++, Sawzall, other

DC

GOOGLE APPSSEARCH

INDEXCRAWLGMAIL...

Architecture

Python. Java. C++

Exterior Network

GWQ

1. Google File System Architecture – GFS II

2. Google Database - Bigtable

3. Google Computation - MapReduce

4. Google Scheduling - GWQ

To meet this workload, typical internet-scale software stack 2003 - 2008

The OS or HW doesn’t do any of the above

Reliability, redundancy all in the “application stack”

Page 50: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Distributed Execution Overview in typical internet-scale workflow

UserProgram

Worker

Worker

Master

Worker

Worker

Worker

fork fork fork

assignmap

assignreduce

readlocalwrite

remoteread,sort

OutputFile 0

OutputFile 1

write

Split 0Split 1Split 2

Input Data

10s of thousands of servers

Technologies such as Hadoop and MapReduce

Page 51: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Internet-scale IT infrastructure

Inp

ut

from

th

e I

nte

rnet

You

r cu

sto

mers

End Result:

Each red block is an inexpensive server = plenty of power for its

portion of workflow

Page 52: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Warehouse Scale Computer programmer productivity framework example

Hadoop– Overall name of software

stack

HDFS– Hadoop Distributed File

System

MapReduce– Software compute framework

• Map = queries • Reduce=aggregates

answers

Hive– Hadoop-based data

warehouse

Pig– Hadoop-based language

Hbase– Non-relationship database

fast lookups

Flume– Populate Hadoop with data

Oozie– Workflow processing

system

Whirr– Libraries to spin up Hadoop

on Amazon EC2, Rackspace, etc.

Avro– Data serialization

Mahout– Data mining

Sqoop– Connectivity to non-Hadoop

data stores

BigTop– Packaging / interop of all

Hadoop components

http://wikibon.org/wiki/v/Big_Data:_Hadoop%2C_Business_Analytics_and_Beyond

Page 53: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

So the real question is:

If we run these immense scale Internet-style workloads

And: – The Internet-sized workload is too large for even maximum size tightly-clustered

memory-shared SMP

– Therefore, workload runs on clusters of 1000s, 10000s of servers • With their corresponding PBs storage, network, power, cooling, software

Given this workload, what is most cost-effective hardware?

We compare many high-end servers vs. thousands of commodity servers

This is the REAL questionFor Internet Scale Data Centers

Page 54: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

TPC-C Benchmark: High-end SMP vs. low-end PC-class server

Low-end server TPC-C is 4x less expensive If we exclude storage costs, low-end server advantage jumps to 12x cheaper. This is meaningful.

4xless

12xless

“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, table 3-1, p. 32 Barroso, Holzle

http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

(from late

2007)

Page 55: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Then, compare Execution Time Parallel Tasks at 3 levels of communication intensity

“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 3-1, p.34 Barroso, Holzle

http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Traditional IT high inter-communication workload = high end SMP has high

inter-process overhead

So what would happen if we increased number

of nodes 130x?

Internet light intercommunicationworkload = small performance

degradation

Past 8 nodes, little additional penalty for increasing cluster size

Page 56: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Performance difference: Internet workload high-end vs. low-end server

“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 3-2, p.35 Barroso, Holzle

http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Note how quickly performance advantage of high-end SMP

diminishes as cluster size increases

At > 2000 cores, 512 low-end servers is within 5%

of 16 high-end servers

12x cost savings at 5% difference

Bottom line: whenever Internet workload is involved (which is too large for any single high-end server

cluster)we do need to think differently about it

That’s why commodity class servers used for

light communication Internet-scale workloads

Internet workload

Page 57: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Bottom line for Internet Scale Workloads:

It makes sense to use consumer grade servers, storage

For Internet-style workloads at Internet scale

It makes sense to use high performance tightly coupled high-end servers

If your workload has high inter-process communication(typical of traditional IT applications)

Page 58: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Therefore, internet workload purpose-built server…. with onboard UPS

Huh?

Why an onboard UPS?

We’ll examine that next

Energy in the form of a UPS on each server

is deployed

As part of strategy to address biggest data

center construction costs

Much more than power outage. Goal is support temporary > 100% power

provisioning in data center

To ride through renewable energy lulls (lack of wind, lack of

solar)

Credit for these ideas: Google 2011 June talk by Luis Andre Barroso at 2011 Federated Computing Research Conference San Jose, Calif

Page 59: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Let’s now examine the

warehouse-level data center

design itself

Ask yourself:What’s biggest cost-savings element?

Page 60: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Internet Scale data center power components…

Image courtesy of DLB Associates: D. Dyer, “Current trends/challenges in datacenter thermal management—a facilities perspective,”presentation at ITHERM, San Diego, CA, June 1, 2006.“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 4-1, p.40 Barroso, Holzle

http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Page 61: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Breakdown of data center energy overheads

Image courtesy of ASHRAE “The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 5-2, p.49 Barroso, Holzlehttp://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Chiller alone is 33% of the cost

UPS alone is 18% of

construction cost

Physical cooling, UPS dominates the electrical power cost

Page 62: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

construction cost of Internet Scale Data Center is Power / Cooling

Facebook’s North Carolina Data Center Goes Live

Facebook: Lulea, Sweden - 290K sq ft (27K sq meters) by late 2012

Facebook – Prinville, Oregon

Has spent $1B on it’s data centers

Open Compute Project

? Reducing power profile reduces

construction cost

Page 63: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Wow. Given that fact…..

Whose data centers are most power efficient?

Reducing power profile = lowers initial CAPEX SIGNIFICANTLY

Therefore, fundamental Internet Scale Data Center goal is:

Decrease Power Usage Effectiveness (PUE)

PUE =

http://gigaom.com/cloud/whose-data-centers-are-more-efficient-facebooks-or-googles/

Total Building Power consumed---------------------------------------------

IT power consumed

Page 64: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google claims its data centers use 50% less energy than competitors

Power Usage Effectiveness– PUE=1.14 means power overhead is

only 14%– Industry average is around 1.8

http://venturebeat.com/2012/03/26/google-data-centers-use-less-energy/

Industry average PUE is about 1.8

http://www.datacenterknowledge.com/archives/2011/05/10/uptime-institute-the-average-pue-is-1-8/

Page 65: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Container modular data center: solving the Power Density issue

Image courtesy of DLB Associates: D. Dyer, “Current trends/challenges in datacenter thermal management—a facilities perspective,”presentation at ITHERM, San Diego, CA, June 1, 2006.

“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 4-2, p.42 Barroso, Holzlehttp://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Without specialized plenums or containerized enclosures, maximum power density of 150-200W / square foot

Due to limits to how much air regular fans can push

Data center can only operate a few minutes without cooling air

Page 66: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Modular Data Center

Value isn’t just time to delivery / flexibility

It’s also Higher Power density = lower construction cost

http://www.youtube.com/watch?v=zRwPSFpLX8I

Page 67: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

That’s why you see such a big modern push on Container Data Centers:

7. PHOENIX ONE, PHOENIX, ARIZ. 538,000 SF

5. MICROSOFT CHICAGO DATA CENTER, Chicago 700,000 SF 2. QTS METRO DATA CENTER, ATLANTA, 990,000 SF

Microsoft’s Chicago Container Data Center

Page 68: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

State of the Modular Data Center

Cyrus One 1 million sq ft “Massively Modular” data center under construction in Phoenix, Arizona

I/O Modular Data Center Assembly line

http://www.datacenterknowledge.com/archives/2012/05/17/cyrusone-going-massively-modular-in-phoenix/

http://www.datacenterknowledge.com/archives/2012/02/06/the-state-of-the-modular-data-center/

http://www.datacenterknowledge.com/archives/2012/01/30/inside-ios-modular-data-center-assembly-line/

Mismatch between rapid workload churn vs. 10+ year data center lifespan = modular data center characteristics strategic possibilities for

new build data centers

Page 69: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

So given all of this

How do I put it all together

In a Warehouse Scale Computer?

Page 70: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google’s Machinery as result of all these factors:

70

Page 71: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Architectural view of Google server and storage hierarchy

71

Page 72: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Clusters through the years

“Google” Circa 1997 (google.stanford.edu) Google (circa 1999)

72

Page 73: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google Data Center (Circa 2000)

Clusters through the years

Google (new data center 2001)

3 days later 73

Page 74: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Recent Google Design

• In-house rack design• PC-class motherboards• Low-end storage and networking hardware• Linux• + in-house software

74

Page 75: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Container Datacenter

75

Run container hotter than normal

human comfort temperature =

big cost savings

Page 76: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google Container Datacenter

76

Page 77: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google: The Dalles, Oregon internet scale data center

77Google Data Center – The Dalles, Oregon

Page 78: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google Data Centers

in 2008:

Page 79: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Google Data Center CAPEX worldwide

Capital expenditures on datacenters:– 1Q12: USD$ 607M– 2011: USD$ 3.4B– 2010: USD$ 4.0B– 2009: USD$ 809M

Each data center between $200M and

$600M

The Dalles, Oregon

Page 80: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

And that…. is what today’s Internet Scale Data Center looks like

Page 81: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

What will a European version of Internet Scale Cloud look like?

Data protection situation still evolving

Europe is Europe– Languages, culture, currencies

Cloud adoption will be very different country to country

Regardless, interest is Hot, Hot, Hot

– 2012: 73% of companies moving to some sort of cloud

– 2012: 55% moving to a private cloud

I believe Europe will adopt best of what’s already been done elsewhere

– In a uniquely European flavor

http://gigaom.com/cloud/will-there-be-an-amazon-of-europe/ http://gigaom.com/cloud/ec-cloud-plan-addresses-data-protection-problem-sort-of/

http://gigaom.com/cloud/5-things-you-need-to-know-about-cloud-in-europe/

http://gigaom.com/europe/dont-worry-europe-youre-about-to-get-a-new-beginning/

Page 82: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Apply lessons from today to Traditional IT as best possible

Source: Egan Ford, IBM Distinguished Engineer, OpenStack presentation: http://xmission.com/~egan/cloud/Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Page 83: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Read all about it, Google published this information into the public domain in 2009

By Google:– Luiz Andre Barroso– Uri Holze

Available to all, free of charge

Download a copy at: http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Video of Luis giving one of these lectures: http://inst-tech.engin.umich.edu/leccap/view/cse-dls-08/4903

http://www.barroso.org/

Page 84: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Let’s review our plans

To meld / meet / build readiness

Page 85: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

To successfully co-exist / thrive with new generation workloads

Understand the Internet Scale Data Center / workload environment / lessons– End to end discovery, monitoring,

operational automation– Differentiate between traditional IT

and internet-scale workloads• For these two categories, architect IT systems

accordingly– Essential role of power efficiency on

CAPEX for new data center costs

$ /

serv

er

TraditionalIT

DataWarehouse

BigData

Internetscale

Views new generation as a

powerful partner Traditional IT workloads

Internet scale warehouse computer

Views the traditional IT as

a powerful enabler

Understand and innovate using these principles within your environment:– Be viewed as a powerful partner and enabler

of these future directions– Architect now, how you wish to your platform,

people, and infrastructure to grow along these lines

– Begin evolving and building skills now

Review attached learning points

Page 86: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Today, our users all have many non-traditional IT alternatives

Traditional IT:

Traditional IT platforms

Traditional IT vendors

Non-traditional alternatives: – The Cloud, the Developing World

What will the effect be on your IT infrastructure, skill

sets, and business models?

Page 87: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Other observations

Think larger than technology

Watch the business models, learn and apply

My additional presentation: “ Disruptive Innovation in Modern IT World”

– http://www.slideshare.net/johnsing1/a-india-csii2012disruptiveinnovationinthemodernitworldv3plenarypresentation

Keeping up with it all:– Necessary today: first thing every day, 1 hour of

industry study, to keep up

– Then share via your own digital footprint• A job skill necessity for today’s world

– Social network, personal exploitation of modern smart devices and tools

– See appendix for resources

Endless possibilities!!

I believe you would know better than I where to apply yourself

$ /

ser

ver

TraditionalIT

DataWarehouse

BigData

Internetscale

Current IT

processor req’d linear

with workload

Internet scale, warehouse scale

computer

New gen workloads

Exascale datacenter

Massive parallelism Flexible system optimization

Workload Optimized

Systems

Page 88: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Summary Computation is inevitably moving into Warehouse Scale Computing supporting The Cloud

– IT Architects, now and in near future, must be aware of and capable of exploiting Internet Data Center Design and Workload experiences to best design the systems of the future

– When the workload is true internet scale, will require the physical and economic mechanisms at play in a Warehouse Scale Computer

This session is the author’s research compilation, Great thanks to Google for the seminal work on which this lecture is based.Download a copy at: http://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Final comments– While WSC’s at one level are simple: a few ten thousand servers and a

LAN……..

– In reality, building cost-effective massive warehouse scale computing platform with necessary reliability, programming productivity, energy cost effectiveness is as difficult and as exciting / stimulating opportunity as IT has ever seen.

– The authors of “Intro to Warehouse Scale Computing” hope that this information will stimulate IT staff and scientists to understand this new area

– In the years to come, our collective understanding / efforts will solve and expand the many fascinating problems and benefits to humanity, arising from warehouse scale computer systems.

Page 89: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Evaluations and chart downloads are online

http://www.ibmtechu.com/hu

sGE06

Page 90: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Together, let’s build a Smarter Planet

Page 91: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning Points

Page 92: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning Points - 1

Rising bandwidth worldwide enables web services delivery model (“The Cloud”)

The Cloud runs in massive data centers with well-connected commodity processors, storage

– With homogenous applications amortized across internet scale user population

These data centers are a different class of large-scale computing machines called:– Warehouse scale computers (WSC)– With huge PB data volumes– Running the easily parallelized, high workload churn, homogeneous platform, fault-

tolerant clustered software stack

Understanding this class of machines:– Important as multi-core processor chip advancement within just a few years– Will make even modest-sized computing systems– Approach the behavior of today’s Warehouse Scale Computer

Page 93: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning points - 2

Building block of choice for WSC is:– Commodity server-class processors, consumer/enterprise grade disk drives,

Ethernet-based networking– Because the internet workload characteristics include easy parallelization

Fault-tolerant software stack mitigates continuous failure rate– Of 10,000s / 100,000s of hardware and software components in WSC– Programmer software stack provides the tools to cost-effectively, time-effectively

program this highly clustered environment– Redundancy in application-level software eliminates need for redundancy in OS or

storage

Software development differs from traditional IT:– Ample parallelism:

• Internet users have a high degree independence from each other– High workload churn:

• Release cycle measured in days and weeks – Platform homogeneity:

• Single organization owns / has technical capability / runs WSC end to end– Fault-tolerant software:

• Makes feasible continuous recovery mode operation of servers / components w/o user application impact

Page 94: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning Points – 3 – Economics / Cost

80% of construction cost of data center due to amount of power and cooling required

Maximizing Power Usage Efficiency is therefore paramount– To reducing capital expenditure as well as operating expenditure– Target PUE = 1.2

Modular Container Data Center architecture is popular:– Mainly because it increases the Power Profile / Power Density– Which in turn significantly reduces the data center construction cost– In addition provides flexibility, much faster time to delivery– Finally, is important tool to help address mismatch between Internet-scale hyper-

workload churn and 10+ year data center lifespan– Modular Container Data Center architecture has considerable merit for any organization

with scale requirements

Page 95: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning Points – 4 – Key Challenges / Opportunities Ahead

Rapidly changing workload– New applications / innovations gain popularity at very fast pace – Often exhibit disruptive workload characteristics (YouTube example)– WSC architecture, container data centers are best practices to cost-effectively best

position / adapt the organization– To disruptive business innovations over the 10+ year lifespan of physical WSC structure

Building balanced systems from imbalanced components– Multi-core processors continue to get faster, become more energy efficient– Memory, disk storage, networking gear not evolving at same pace in either performance

or energy efficiency– Research / innovation must shift to these subsystems else further increases in processor

power will not be able to provide further WSC improvement

Page 96: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning Points – 5

Curbing energy cost– We must continue to find ways to ensure performance improvements are

accompanied by corresponding energy efficiency improvements– Otherwise scarce future energy budget will increasingly curb growth in computing

capabilities

Internet-style workloads– Future performance gains will be delivered by more multi-cores, not clock speed– Future large scale systems will continue to increasingly exhibit characteristics of

today’s Internet Scale Data Centers and Workloads

Page 97: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Supplemental Resources

Page 98: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Energy Proportionality

Page 99: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Activity profile 5,000 Google servers over period of 6 months

“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 5-4, p.55 Barroso, Holzlehttp://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Majority of time, server utilization is in 10-50% range

Obviously some opportunity to increase processor util. %

The real question: how much power / cooling did I have to pay for in this data center to run these idling servers?

Page 100: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

SPECpower_ssj2008: traditional IT servers consume nearly 70% power even when idling!

“The Data Center as a Computer: Introduction to Warehouse Scale Computing”, figure 5-3, p.53 Barroso, Holzlehttp://www.morganclaypool.com/doi/pdf/10.2200/S00193ED1V01Y200905CAC006

Server consumes 68% of peak power

requirement when idling!

i.e. at a 30% utilization, we’re using 75% of max power

A lot!!!! Even though most of server time < 50% utiliz. I’m paying 70% energy cost

Page 101: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Energy Proportionality

2011 servers have gotten a lot better

By closing this gap,construction costs of internet

scale data center feasible

This gap = excessive data center construction

cost

Enables warehouse

scale computer to be affordable

Credit for all these ideas is to Google 2011 June talk by Luis Andre Barroso at 2011 Federated Computing Research Conference San Jose, Calif

Page 102: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

But to fully exploit Energy Proportionality…..

Rest of IT infrastructure needs to catch up

Servers today: 3x– Have improved greatly since 2008

But: – Currently little/no focus on energy proportionality in:

Networking equip: 1.2x

Storage: 1.3x– Hard to do because we’re spinning the disk drive constantly – Spinning drives -> flash?

Dynamic RangeBigger is better

Means uses nearly same power whether it’s idle or fully utilized

Affects data center construction costs

Page 103: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Is Internet Scale = High Performance Computing (HPC)? No

HPC clusters:1. Recovery model: OK to pause entire job,

restart computation from checkpoint

2. Requires significant CPU supporting– large numbers of synchronized tasks – which intensely communicate

3. Typically single binary, single job on 1000s of nodes

4. CPUs may run at 100% for days/weeks

5. Building block of choice: high perf, high avail high-end SMPs with high shared memory interconnect bandwidth for intense inter-process communication

Warehouse scale computers:1. Recovery model: gracefully tolerate large

#s component faults– operating near-continuous recovery

state

2. Requires significant CPU but individual tasks less synchronized– Little or no inter-communications– Internet workload = ample parallelism

3. Diverse set of applications– Hyper-pace workload churn / release

cycles

4. CPU utilization varies, rarely 90% due to need to reserve capacity for Internet spikes or to cover failed cluster components

5. Building blocks of choice: commodity server-class machines with direct attached disk drives, Ethernet-based interconnect

Page 104: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Resources for your Internet Scale Workload and Data

Centerjourney

Page 105: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

How to get ahead and thrive in this new world?

2012: devote 1st hour of day to keeping current

–No longer optional

Establish power-knowledge digital footprint, intelligently sharing what you find

–Don’t email what you find (too much email already)

–Use social networking, social bookmarking, blogs, etc

Become a power user of your smartphone’s ecosystem

Page 106: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

KeepingCurrent

UsingJohn Sing’sbookmarks

You may use me as one source of ‘who/what to follow’– Connect with me: http://www.linkedin.com/in/johnsing

External: my daily list of social bookmarks is: – http://delicious.com/atsf_arizona

IBM colleagues may see my IBM Intranet webpage:– http://snjgsa.ibm.com/~singj/

IBM Colleagues can see my IBM SONAS intranet web page

– http://snjgsa.ibm.com/~singj/public/sonas_index.html [email protected]

Page 107: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

See video reviewing the Space-Time-Travel example by IBM Distinguished Engineer (SWG) on Big Data – superb insight into Big Data

http://gigaom.com/2011/03/23/jeff-jonas-ibm/

Jeff Jonas/Las Vegas/IBM

Page 108: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Keeping current – More places to connect

Find out what your colleagues are doing

https://www.facebook.com/pages/IBM-NAS/156301741086498

https://www.facebook.com/IBMRedbooks

https://www.facebook.com/peopleforasmarterplanet

http://storagecommunity.org/

https://www.ibm.com/developerworks/mydeveloperworks/blogs/InsideSystemStorage/?lang=en

Page 109: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learn what’s out there: McKinsey Global Report on Big Data

A seminal work on this fast evolving technology, critically important technology.

While 153 pages long - if you understand the content of this presentation and realize that Big Data is insanely important to future IT viability skills....

This paper gives superb, concrete, well-substantiated ideas on what Big Data is being used for today, as we speak, to create the business models of the future

You may download a copy here:

http://www.mckinsey.com/mgi/publications/big_data/index.asp

Page 110: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Forbes Sept 2011: Impact of Social Media on Corporate

http://www.forbes.com/sites/techonomy/2011/09/07/social-power-and-the-coming-corporate-revolution/

Page 111: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://www.theregister.co.uk/hardware/storage/

Page 112: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://gigaom.com/

Page 113: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://www.datacenterknowledge.com/special-report-the-worlds-largest-data-centers/

Develop your list of daily reading and updating…

Page 114: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Here’s one view of world’s largest data center:

Questions:

Do you know where the largest data centers are?

Arew we tracking what they do, and why?

We could, we should!

http://www.datacenterknowledge.com/special-report-the-worlds-largest-data-centers/worlds-largest-data-center-350-e-cermak/

Page 115: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://www.datacenterknowledge.com/archives/2009/05/14/whos-got-the-most-web-servers/

Page 116: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

All about the Hadoop Distributed File System (open source)

http://hadoop.apache.org/hdfs/

Page 117: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://www.highscalability.com

Of particular interest is the “Real Life Architectures” tab

Page 118: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Hugely important, keep inspiring ourselves – one of my favorites:

http://www.ted.com/ - superb world class non-profit dedicated to Ideas Worth Spreading in technology, entertainment, design

Page 119: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

http://www.slideshare.net

Search on the topic that you’re researching

– Competitors in particular

Find fantastic number of downloadable presentations

– Some better than others, but quickly, you’ll learn to sift find great quality for yourself

Page 120: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Recommend you download, read,this very informative IBM book

“Understanding Big Data” – Published April 2012– Free download– Well worth reading to understand components

of Big Data, and how to exploit

Part 1: The Big Deal about Big Data– Chapter 1 – What is Big Data? Hint: You’re a

Part of it Every Day– Chapter 2 – Why Big Data is Important– Chapter 3 – Why IBM for Big Data

Part II: Big Data: From the Technology Perspective

– Chapter 4  - All About Hadoop: The Big Data Lingo Chapter

– Chapter 5 – IBM InfoSphere Big Insights – Analytics for “At Rest” Big Data

– Chapter 6 – IBM InfoSphere Streams – Analytics for “In Motion” Big Data

http://public.dhe.ibm.com/common/ssi/ecm/en/iml14297usen/IML14297USEN.PDFDownload your free copy here

Page 121: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Interested in reading more about Competitive Advantage Analytics-based applications? Easy-to-read pages in this IBM book:

Download it (3.8 MB Acrobat Reader file) at:

ftp://ftp.software.ibm.com/common/ssi/pm/bk/n/imm14055usen/IMM14055USEN.PDF

User Interface LayerDashboards, Mashups, Search,Ad hoc reporting, Spreadsheets

Analytic Process LayerReal-time computing and analysis, stream computing,

entity analytics, data mining, content management, text analytics, etc.

Infrastructure layerVirtualization, central end to end management, control,

data proximity, deployment on global virtual file server with geographically dispersed storage

Secu

rityau

tho

rization

Location ofcustomer

competitive advantage applications

This book defines everything you need to know about Competitive Advantage modern analytics applications. Interesting reading.

If you are needing a quick overview of modern Analytics IT capability, start on page 14, read through page 48.

Page 122: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

IBM Smarter Planet Big Data website

http://www-03.ibm.com/systems/data/flash/smartercomputing/bigdata.html

Page 123: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

IBM Software Group Big Data web site

http://www-01.ibm.com/software/data/bigdata/

Page 124: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning Points: Ten Big Data Realities

Here are the first ten points that I want you think about when you’re grokking big data:

Oracle is not big data

Big data is not traditional Relational Database Management System (RDBMS)

Big data is not a Exadata

Big data is not a EMC VMAX

Big data is not highly structured

Big data is not centralized

IT people are not driving big data initiatives

Big data is not a pipe dream – big data initiatives are adding consumer and business value today. Right now. Every second of every minute of every hour of every day.

Big data has meaning to the enterprise

Data is the next source of competitive advantage in the technology business.

Source: David Vellante, 1Q2011

Source: Wikibon.org, March 1, 2011 public broadcat on “Big Data”, http://wikibon.org/blog/ten-%E2%80%9Cbig-data%E2%80%9D-realities-and-what-they-mean-to-you/

Page 125: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Learning points: What does Big Data mean to IT infrastructure professionals?

Big data means the amount of data you’re working with today will look trivial within five years.

Huge amounts of data will be kept longer and have way more value than today’s archived data.

Business people will covet a new breed of alpha geeks. You will need new skills around data science, new types of programming, more math and statistics skills and data hackers…lots of data hackers.

You are going to have to develop new techniques to access, secure, move, analyze, process, visualize and enhance data; in near real time.

You will be minimizing data movement wherever possible by moving function to the data instead of data to function. You will be leveraging or inventing specialized capabilities to do certain types of processing- e.g.early recognition of images or content types – so you can do some processing close to the head.

The cloud will become the compute and storage platform for big data which will be populated by mobile devices and social networks.

Metadata management will become increasingly important.

You will have opportunities to separate data from applications and create new data products.

You will need orders of magnitude cheaper infrastructure that emphasizes bandwidth, not IOPs - and data movement with efficient metadata management.

You will realize sooner or later that data and your ability to exploit it is going to change your business, social and personal life; permanently.

Source: David Vellante, 1Q2011

Source: Wikibon.org, March 1, 2011 public broadcat on “Big Data”, http://wikibon.org/blog/ten-%E2%80%9Cbig-data%E2%80%9D-realities-and-what-they-mean-to-you/

Page 126: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

More information for my IBM colleagues - read transcript of Big Data Overview

http://snjgsa.ibm.com/~singj/public/2011_Big_Data_Modern_Analytics_Tutorial/

Page 127: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

More information for my IBM colleagues

Below is the 2012 IBM Research Global Technology Outlook:– https://w3-connections.ibm.com/wikis/home?lang=en_US#/wiki/Wd99c91e6c090_42d6_

bbef_095a93a1bc63

Below is the IBM Research Global Technology Outlook 2011 Overview which includes our first discussions of Big Data:

– http://snjgsa.ibm.com/~singj/public/2011_Prague_IBM_Systems_Conference/STG%20Tech%20Conference%20GTO%202011%20from%20Dr%20Matthias%20Kaiserwerth.ppt

See all the IBM IBM Research Global Technology Outlook 2011 charts at: – http://w3.ibm.com/articles/workingknowledge/2010/12/res_gto_2011.html

Page 128: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Internet Scale Data Centers A different scale and set of server storage and facility

economics

Implies where our own strategies, skill sets, and architectures can expand:

– With additional styles of thinking, architecture– If you think 2012 is growing fast– Going to take off even more in 2013– Mucho resources in appendix to these charts

We are all at an inflection point

Growth areas

Traditional areas

Page 129: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Thank You

Page 130: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Appendix

Following are charts from Budapest session xCL01, from Egan Ford, IBM Distinguised Engineer, System X / STG Cloud Strategy

It is his research on this similar topic of Internet Scale Workloads and Data Center Design

Page 131: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Today: two different types of IT

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Internet scale wkloadsTransactional IT

Page 132: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Today’s two major IT workload types

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/ Transactional IT Internet scale wkloads

Page 133: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

How to build these two different clouds

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Transactional ITInternet scale wkloads

Page 134: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

What You (Consumer) Get with These Clouds:

Source: http://it20.info/2012/02/the-cloud-magic-rectangle-tm/

Page 135: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Policy-based Clouds and Design-for-fail Clouds are purpose optimized Infrastructure Management solutions

Policy-based Clouds

• Purpose optimized for longer-lived virtual machines managed by Server Administrator

• Centralizes enterprise server virtualization administration tasks

• High degree of flexibility designed to accommodate virtualization all workloads

• Significant focus on managing availability and QoS for long-lived workloads with level of isolation

• Characteristics derived from exploiting enterprise class hardware

• Legacy applications

Design-for-fail Clouds

• Purpose optimized for shorter-term virtual machines managed via end-user or automated process

• Decentralized control, embraces eventual consistency, focus on making “good enough” decisions

• High degree of standardization

• Significant focus on ensuring availability of control plane

• Characteristics driven by software

• New applications

Page 136: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Some OpenStack Public Use Cases

• Internap– http://www.internap.com/press-release/internap-announces-world%E2%80%99s-first-

commercially-available-openstack-cloud-compute-service/

• Rackspace Cloud Servers, Powered by OpenStack– http://www.rackspace.com/blog/rackspace-cloud-servers-powered-by-openstack-beta/

• Deutsche Telekom– http://www.telekom.com/media/media-kits/104982

• AT&T– http://arstechnica.com/business/news/2012/01/att-joins-openstack-as-it-launches-cloud-

for-developers.ars

• MercadoLibre– http://openstack.org/user-stories/mercadolibre-inc/mercadolibre-s-bid-for-cloud-

automation/

• NeCTAR– http://nectar.org.au/

• San Diego Supercomputing Center– http://openstack.org/user-stories/sdsc/

Page 137: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

OpenStack design tenets focus on delivering essential infrastructure on an available, scalable, elastic control plane

Sources:http://www.openstack.org/downloads/openstack-compute-datasheet.pdfhttp://wiki.openstack.org/BasicDesignTenets

Basic Design Tenets

1) Scalability and elasticity are our main goals

2) Any feature that limits our main goals must be optional

3) Everything should be asynchronous. If you can't do something asynchronously, see #2

4) All required components must be horizontally scalable

5) Always use shared nothing architecture (SN) or sharding. If you can't Share nothing/shard, see #2

6) Distribute everything. Especially logic. Move logic to where state naturally exists.

7) Accept eventual consistency and use it where it is appropriate.

8) Test everything. We require tests with submitted code. (We will help you if you need it)

OpenStack Leadership's vision statement

“essential Infrastructure, support platform”

Page 138: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

OpenStack

Source: http://ken.pepple.info/openstack/2012/09/25/openstack-folsom-architecture/

Page 139: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

OpenStack is comprised of seven core projects that form a complete IaaS solution

Compute (Nova)

Storage (Cinder)

Network (Quantum)

Provision and manage virtual resources

Dashboard (Horizon)Self-service portal

Image (Glance)Catalog and manage server images

Identity (Keystone)Unified authentication, integrates with existing systems

Object Storage (Swift)petabytes of secure, reliable object storage

IaaS

Source: http://ken.pepple.info/openstack/2012/09/25/openstack-folsom-architecture/

IaaS

Page 140: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Compute delivers a fully featured, redundant, and scalable cloud computing platform

Architecture•

Sources:http://ken.pepple.info/openstack/2012/09/25/openstack-folsom-architecture/http://openstack.org/projects/compute/

Key Capabilities:

•Manage virtualized server resources• CPU/Memory/Disk/Network Interfaces 

•API with rate limiting and authentication

•Distributed and asynchronous architecture• Massively scalable and highly available system

•Live guest migration• Move running guests between physical hosts

•Live VM management (Instance)• Run, reboot, suspend, resize, terminate instances

•Security Groups

•Role Based Access Control (RBAC)• Ensure security by user, role and project

•Projects & Quotas

•VNC Proxy through web browser

Page 141: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Compute management stack control plane is built on queue and database

Key Capabilities:

• Responsible for providing communications hub and managing data persistence

• RabbitMQ is default queue, MySQL DB– Documented HA methods– ZeroMQ implementation available to decentralize

queue

• Single “cell” (1 Queue, 1 Database) typically scales from 500 – 1000 physical machines

– Cells can be rolled up to support larger deployments

• Communications route through queue– API requests are validated and placed on queue– Workers listen to queues based on role or role +

hostname– Responses are dispatched back through queue

Page 142: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

nova-compute manages individual hypervisors and compute nodes

Key Capabilities:

• Responsible for managing all interactions with individual endpoints providing compute resource, e.g.

•-- Attach iSCSI volume to phsyical host, map to guest as additional HDD

• Implementations direct to native hypervisor APIs– Avoids abstraction layers that bring least common

denomination support– Enables easier exploitation of hypervisor

differentiators

• Service instance runs on every physical compute node, helps to minimize failure domain

• Support for security groups that define firewall rules

• Support for– KVM– LXC– VMware ESX/ESXi (4.1 update 1)– Xen (XenServer 5.5, Xen Cloud Platform)– Hyper V

Page 143: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

nova-scheduler allocates virtual resources to physical hardware

Key Capabilities:

• Determines which physical hardware to allocate to a virtual resource

• Default scheduler uses a series of filters to reduce set of applicable hosts and uses costing functions to provide Weight

• Not a focus point for OpenStack– Default implementation finds first fit– Shorter the workload lifespan, less critical the

placement decision

• If default does not work, often deployers have specific requirements and develop custom

Page 144: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

nova-api supports multiple API implementations and is the entry point into the cloud

Key Capabilities:

• APIs supported– OpenStack Compute API (REST-based)

– Similar to RackSpace APIs– EC2 API (subset)

– Can be excluded– Admin API (nova-manage)

• Robust extensions mechanism to add new capabilities

Page 145: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Network automates management of networks and attachments (network connectivity as a service)

Key Capabilities:

•Responsible for managing networks, ports, and attachments on infrastructure for virtual resources

•Create/delete tenant-specific L2 networks

•L3 support (Floating IPs, DHCP, routing)

•Moving to L4 and above in Grizzly

•Attach / Detach host to network

•Similar to dynamic VLAN support

•Support for– Open vSwitch– OpenFlow (NEC & Floodlight controllers)– Cisco Nexus– Niciria

Architecture

Page 146: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Cinder manages block-based storage, enables persistent storage

Key Capabilities:

• Responsible for managing lifecycle of volumes and exposing for attachment

• Structure is a copy of Compute (Nova), sharing same characteristics and structure in API server, scheduler, etc.

• Enables additional attached persistent block storage to virtual machines

• Support for booting virtual machines from nova-volume backed storage

• Allows multiple volumes to be attached per virtual machine

• Supports following– ISCSI– RADOS block devices (e.g. Ceph distributed file

system)– Sheepdog– Zadara

Architecture

Page 147: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Identity service offers unified, project-wide identity, token, service catalog, and policy service designed to integrate with existing systems

Key Capabilities:

• Identity service provides auth credential validation and data about Users, Tenants and Roles

• Token service validates and manages tokens used to authenticate requests after initial credential verification

• Catalog service provides an endpoint registry used for endpoint discovery.

• Policy service provides a rule-based authorization engine and the associated rule management interface.

• Each service configured to serve data from pluggable backend

– Key-Value, SQL, PAM, LDAP, PAM, Templates

• REST-based APIs

Page 148: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Image service provides basic discovery, registration, and delivery services for virtual disk images

Key Capabilities:

• Think Image Registry, not Image Repository

• REST-based APIs

• Query for information on public and private disk images

• Register new disk images

• Disk images can be stored in and delivered from a variety of stores (e.g. SoNFS, Swift)

• Supported formats– Raw– Machine (a.k.a. AMI)– VHD (Hyper-V)– VDI (VirtualBox)– qcow2 (Qemu/KVM)– VMDK (VMWare)– OVF (VMWare, others)

Referenceshttp://openstack.org/projects/image-service/

Page 149: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

Dashboard enables administrators and users to access and provision cloud-based resources through a self-service portal

Key Capabilities:

• Thin wrapper over APIs, no local state

• Registration pattern for applications to hook into

• Ships with three central dashboards, a “User Dashboard”, a “System Dashboard”, and a “Settings

• Out-of-the-box support for all core OpenStack projects– Nova, Glace, Switch, Quantum

• Anyone can add a new component as a “first-class citizen”.– Follow design and style guide.

• Visual and interaction paradigms are maintained throughout.

• Console Access

Referenceshttp://horizon.openstack.org/intro.html

Page 150: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

OpenStack Resources

• Forums– http://forums.openstack.org/

• Wiki– http://wiki.openstack.org/

• Documentation– http://docs.openstack.org/

• Mailing Lists– http://wiki.openstack.org/MailingLists

• OpenStack Project Management– https://launchpad.net/openstack

• Blogs– http://planet.openstack.org

• Real-time chat room– #openstack and #openstack-dev on irc://freenode.net (443 users currently logged in)

• Rackspace Reference Architectures– http://www.referencearchitecture.org/

• Easy Install– http://www.hastexo.com/resources/docs/installing-openstack-essex-20121-ubuntu-1204-precise-

pangolin

Page 151: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

IBM Resources/Solutions for OpenStack Available Today

• developerWorks– https://www.ibm.com/developerworks/mydeveloperworks/wikis/home?lang=en#/wiki/

OpenStack– Google: openstack IBM developerworks

• xCAT (FOSS) for 0-day deployment– xCAT OpenStack Paper (CATStack)– Automated qcow2 image creation for Glance– HW control– Bare-metal discovery and bring up

•Firmware, Base OS, etc…

• IBM Intelligent Cluster Solutions (see Matt Ziegler's PPT)– Preconfigured Switches– Rack and stacked and ready to go– Lab Services for 0-day

Page 152: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

IBM Resources/Solutions for OpenStack Available Today

• All IBM System Software and Tools can coexist with OpenStack.– Director, ASU, lflash, etc…

• SoNAS for shared file (NFS, SMB)

• XIV for block storage (Nova Volume)

• iDPX for scale-out Nova Compute and Swift

• BNT switches for OpenFlow and Quantum

• GPFS for iSCSI/block (Nova Volume) or file.

Page 153: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

OpenStack Demo Setup

10.0.9.10 10.0.9.11 10.0.9.12 10.0.9.13 10.0.9.X

172.20.249.10 172.20.249.11 172.20.249.12 172.20.249.13 172.20.249.X

os-essex0 os-essex1 os-essex2 os-essex3 os-essexX

Control Nodes Compute Nodes

Private Networks: eth0: 172.20.249/24 vm: 172.20.250/24

Public Networks: eth1: 10.0.9.0/25 vm: 10.0.9.128/25

computenetwork

computenetwork

computenetwork

computenetworkschedulervolumeconsoleglanceapi

computenetworkschedulervolumeconsoleglanceapi

Scale OutHA Active/Passive

VMVM

VMVM

VMVM

VMVM

VM Firewall

Page 154: To_Infinity_and_Beyond_Internet_Scale_Workloads_Data_Center_Design_v6

© 2012 IBM Corporation

IBM Systems Technical Universities– Budapest, Hungary – October 15-19sGE06

PPT’s and Videos: http://xmission.com/~egan/cloud/