Next Generation Cloud Computing: Advanced Services,

Architecture, and Technologies

Joe Mambretti, Director, (j-mambretti@northwestern.edu)

International Center for Advanced Internet Research (www.icair.org)

Northwestern University

Director, Metropolitan Research and Education Network (www.mren.org)

Partner, StarLight/STAR TAP, PI-OMNINet (www.icair.org/omninet)

Technische Universit Carolo-Wilhelmina

zu Braunschweig

Braunschweig,

July 1-3, 2009

• Creation and Early Implementation of Advanced Networking Technologies - The Next Generation Internet All Optical Networks, Terascale Networks, Networks for Petascale Science

• Advanced Applications, Middleware, Large-Scale Infrastructure, NG Optical Networks and Testbeds, Public Policy Studies and Forums Related to NG Networks

• Three Major Areas of Activity: a) Basic Research b) Design and Implementation of Prototypes c) Operations of Specialized Communication Facilities (e.g., StarLight)

Accelerating Leading Edge Innovation

and Enhanced Global Communications

through Advanced Internet Technologies,

in Partnership with the Global Community

Introduction to iCAIR:

Invisible Nodes,

Elements,

Hierarchical,

Centrally Controlled,

Fairly Static

Traditional Provider Services:

Invisible, Static Resources,

Centralized Management,

Highly Layered

Distributed Programmable Resources,

Dynamic Services,

Visible & Accessible Resources,

Integrated As Required, Non-Layered

Limited Services, Functionality,

Flexibility

Unlimited Services, Functionality,

Flexibility

Paradigm Shift – Ubiquitous Services Based on Large Scale

Distributed Facility vs Isolated Services Based on Separate

Component Resources

A Next Generation Architecture: Distributed Facility

Enabling Many Types Network/Services

Commodity

Internet

Environment: VO

Environment:

International Gaming Fabric

Environment:

Control Plane

TransLight Environment: Real Org

Environment: Sensors

Environment: Intelligent

Power Grid Control

Environment: Real Org1

Environment:

Large Scale System Control

Environment: Global App

Environment: Financial Org

Environment: Gov AgencyEnvironment: RFIDNet

Environment: Bio Org

Environment: Lab

Environment: Real Org2

SensorNetFinancialNet

HPCNet

MediaGridNet

R&DNet

RFIDNet

BioNet

PrivNet

GovNet1

MedNet

Cloud Context (1)

• In General, Clouds Are A Means To Support Large Scale Computing

and Data Capabilities for Distributed On-Demand Resources Using

Data Networks (WANs)

• There Are Many Different Types of Clouds

• Some Are Oriented Toward Services (e.g., Web 2.0 Based)

• Some Are Oriented Toward Resources

– For Example, On-Demand Computing Instances Using

Infrastructure As A Service Techniques (IaaS, Amazon EC2, S3,

etc., Eucalyptus)

• Some provide Large Scale On-Demand Computing Capacity (G

FS/MapReduce/Bigtable, Hadoop, Sector, etc

• Some Support Public Services Provided By Global Corporations,

Some Support Private Enterprises Through External Resources,

Some Support Private Organizations Through Internal Resources –

and There Are Many Variations and Hybrids

5

Cloud Context (2)

• To Date, Clouds Have Been Successful

• However, Current Clouds Have Limitations

• For Example, They Do Not Necessarily Provide

Optimal Performance

• Also, Current Clouds Are Oriented Toward

Providing Support For Many Billions of Small

Data Over Commodity “Best Effort” Routed

Networks

• Current Clouds Do Not Provide Optimal Support for

Large Capacity Data Flows and Extremely Large

Amounts of Individual Data Components

• They Have Not Been Integrated Into Next

Generation Networking Capabilities

• They Do Not Handle Specialized Data Well – e.g.,

Digital Media

Cloud Context (3)

• If Clouds Are Successful, Why Improve Them

Despite Limitations?

• They Are Successful for Today’s Consumer and

Enterprise Services –

• However, They Will Not Meet Future Challenges

Using Current Architectures and Technologies

• Illustration – And Also Motivation For Improving

the State-of-the-Art – Large Scale Science

• Why Large Scale Science?

• Large Scale Science Provides a Looking Glass Into

the Future

A Scientific Perspective

• Scientific Research Requires The Resolution of Extremely Complex Problems

• Scientific Research Requires The Design And Creation Of Specialized Tools

• Increasingly, These Tools Are Being Created Using Digital Technologies

• Because of the Complexity and Scale of Major Scientific Problems, Many Areas of Research Encounter Technical Barriers Years Before They Are Recognized By Other Domains

• Technical Solutions That Are Created Later Migrate To Wider Communities

• Can Large Scale Science Use Clouds? Not Until the Limitations Described Earlier Are Addressed

Motivation: Data-Intensive Science &

Engineering-e-Science Community Resources

ATLAS

Sloan Digital Sky

Survey

LHC

ALMA

Magnetic Fusion Energy

Source: DOE

Source: DOE

New Sources

Of Power

Spallation Neutron Source (SNS) at ORNL

Source: DOE

USGS Images 10,000 Times

More Data than Landsat7

Shane DeGross, Telesis

USGS

Landsat7 Imagery

100 Foot Resolution

Draped on elevation data

New USGS Aerial Imagery

At 6-inch Resolution

Source: EVL, UIC

Today’s Aerial Imaging is >500,000

Times More Detailed than Landsat7

30 meter pixels

4 centimeter pixels

Shane DeGross

Laurie Cooper, SDSU

SDSU Campus

Source: Eric Frost, SDSU

iGrid 2005

UCSD

4K Digital Media

Ultra High Definition

Digital

Communications• NTT, Japan

NTT’s digital communications using

SHD transmits extra-high-quality,

digital, full-color, full motion images.

4k pixels horizontal, 2k vertical

4 * HDTV – 24 * DVD

. www.onlab.ntt.co.jp/en/mn/shd

• Center for Computation and Technology,

Louisiana State University (LSU), USA

• Northwestern University

• MCNC, USA

• NCSA, USA

• Lawrence Berkeley National Laboratory,

USA

• Masaryk University/CESNET, Czech

Republic

• Zuse Institute Berlin, Germany

• Vrije Universiteit, NL

www.cct.lsu.edu/Visualization/iGrid2005

http://sitola.fi.muni.cz/sitola/igrid/

• Interactive visualization coupled with

computing resources and data storage

archives over optical networks enhance

the study of complex problems, such as

the modeling of black holes and other

sources of gravitational waves.

• HD video teleconferencing is used to

stream the generated images in real time

from Baton Rouge to Brno and other

locations

High-Performance Digital MediaFor Interactive Remote Visualization (2006)

OptIPuter JuxtaView Software for Viewing

High Resolution BioImages on Tiled Displays

30 Million Pixel Display

NCMIR Lab UCSDSource: David Lee, Jason

Leigh, EVL, UIC

Components Comprising Environment

• Overall Architecture

• Compute Nodes

• Storage Performance

• Network Architecture, Protocols,

Performance and Technology

Note=>Ultra High Performance Networks

Can Make Remote Data Appear To Be Local

• Proof of Concept – Large Scale

Testbed/Prototype

• Integration With Emerging Technologies, e.g.,

Massive Multicore, FPGAs, Customized

Integrated Components, etc.

Architecture (1)

Storage Compute

Data

Super-Computer Model:

•Expensive

•IO is a bottleneck

Alternative Model:

•Inexpensive,

•Parallel data IO

•Examples: Hadoop

•Sphere/Sector

Source: NCDM, UIC

Architecture (2)

Parallel/Distributed

Programming With MPI,

etc.:

•Flexible and powerful.

•BUT Very Complicated

•No Data Locality

Sector/Sphere Model:

•Very Simple to Apply UDF to

All Data in Parallel;

•Exploits Data Locality

•Limited to Certain Data

Parallel Applications.

Source: NCDM, UIC

What is Sector/Sphere?

• Sector/Sphere = Wide Area Cloud Providing

On-Demand Computing Capacity

• Sector: Distributed Storage System

• Sphere: Run-time Middleware Applies User

Defined Functions (UDF) to Sector Datasets.

• Open Source Software, GPL/LGPL, written in

C++.

• Initiated 2006, Current Version 1.19

• http://sector.sf.net

Source: NCDM, UIC

Sector

• Sector: Provides Long Term Persistent Storage to

Large Datasets Managed as Distributed Indexed Files.

• File Segments Are Placed Throughout Distributed

Storage Managed by Sector.

• Sector Generally Replicates Data To

• Ensure Longevity,

• Decrease the Latency When Retrieving It,

• Provide Opportunities for Parallelism.

• Sector is Designed to Take Advantage of Wide Area

High Performance Networks When Available.

•Sector Can Address Issues Of Extremely Large Data Sets,

Including Very Large Scale Science Data Sets

Source: NCDM, UIC

Sphere

• Sphere: Designed To Execute User Defined Functions

(UDF) In Parallel Using a Stream Processing Pattern for

The Data That Is Managed By Sector

• UDFs Are Applied To Every Data Record In a Data Set

Managed by Sector

• Each Data Segment Is Processed Independently

Providing a Natural Parallelism

• The Sector/Sphere Design Results in Allowing Data

To Be Frequently Processed in Place Without

Moving It

• If Data Must be Moved, It Can Be Transported Over High

Performance Channels With High Performance

Protocols

Source: NCDM, UIC

Comparing Hadoop and Sector

Hadoop Sector

Storage Cloud Block-based file

system

File-based

Programming

Model

MapReduce MapReduce&

UDF

Protocol TCP UDP-based

protocol (UDT)

Replication At time of writing Periodically

Security Within 6 months Security (HIPAA)

Language Java C++

23

Proof of Concept: Large Scale

Testbed/Prototype

• Theories Must Be Proven in Practice Using Real Facilities

• Questions: Can This Concept Scale?

Will It Work At Extreme Scales?

Can High Performance Be Achieved?

• Lab Modeling and Simulation Cannot Substitute for

Real Empirical Studies

• Experimental Research testing Is Required

-- Using Real World Large Scale Facilities

• Distributed Environments and Infrastructure

• With National Science Foundation Funding A Research

Consortium (NCDM, iCAIR, Et Al) Has Created:

• An International Scale TeraFlow Testbed Using NLR and

the Global Lambda Integrated Facility (GLIF)

• A National Scale Open Cloud Testbed, Based on the NLR

Source: NLR

Teraflow 1 & 2 TestbedTeraflow Network is Built Over the

National Lambda Rail and GLIF Seoul

Asia EUGLIF

StarLight – “By Researchers For Researchers”

Abbott Hall, Northwestern University’s

Chicago downtown campusView from StarLight

StarLight is an experimental optical infrastructure andproving ground for network services optimized forhigh-performance applicationsGE+2.5+10GEExchangeSoon:Multiple 10GEsOver Optics –World’s “Largest”10GE ExchangeFirst of a KindEnabling InteroperabilityAt L1, L2, L3

iCAIR: Founding Partner of the Global Lambda Integrated Facility

Available Advanced Customizable Network Resources

Visualization courtesy of Bob Patterson, NCSA; data compilation by Maxine Brown, UIC.

www.glif.is

GLIF is a consortium of institutions, organizations, consortia and country

National Research & Education Networks who voluntarily share optical

networking resources and expertise to develop the Global LambdaGrid for the

advancement of scientific collaboration and discovery.

Enables the Creation of Distributed Virtual Environments

Open Cloud Testbed – 2009

6 Locations

• 8 racks

• 256 Nodes

• 1024 Cores

• 10+ Gb/s

32

MREN

CENIC Dragon

Hadoop

Sector/Sphere

Thrift

Eucalyptus

C-Wave

Example: Sorting a TeraByte

• Data is Split Into Multiple Small Files,

Scattered On All Nodes

• Stage 1: On Each Node, an SPE Scans

Local Files, Sends Each Record To a

“Bucket File” On a Remote Node

According To The key, So That All

Buckets Are Sorted.

• Stage 2: On eEach Destination Node, an

SPE Sorts All Data Inside Each Bucket.

Source: NCDM, UIC

TeraSort Using Sector & Data-Parallel UDFs

on Open Cloud Testbed

10-byte 90-byte

Key Value

10-bit

Bucket-0

Bucket-1

Bucket-1023

0-1023

Stage 1:

Hash based on

the first 10 bits

Bucket-0

Bucket-1

Bucket-1023

Stage 2:

Sort each bucket

on local node

Binary Record 100 bytes

Source: NCDM, UIC

Performance Results: TeraSort on Open

Cloud Testbed

Data Size Sphere Hadoop (3

replicas)

Hadoop (1

replica)

UIC 300GB 1265 2889 2252

UIC +

StarLight

600GB 1361 2896 2617

UIC +

StarLight +

Calit2

900GB 1430 4341 3069

UIC +

StarLight +

Calit2 + JHU

1200GB 1526 6675 3702

Run time: seconds

Sector v1.16 vs Hadoop 0.17

Source: NCDM, UIC

Terasort on Open Cloud Testbed

Source: NCDM, UIC

Testbed Demonstrations With National Science Foundation

at the Annual Conference of

The American Association for the Advancement of Science

February 2009

Using An Optical Fiber Extension from StarLight/GLIF

NSF Director

Former NSF Director NCDM Director

Future: On-Going Expansion

• Expansion Using More Resources

• Additional Enhancements

• Integration With Emerging Technologies

• Expansion Across National Fabrics

• Expansion Across International Fabrics,

Using GLIF, StarLight

• Additional Communities

For More Information

• iCAIR: www.icair.org

• NCDM: www.ncdm.uic.edu

• Open Cloud Testbed:

www.opencloudconsortium.org

• Sector: sector.sourceforget.net