HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress...

101
© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. HPC in the Cloud Keynote, IEEE 8 th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories [email protected] Work with Nigel Cook, Paolo Faraboschi, Abhishek Gupta, Laxmikant V. Kalé, Sudarsun Kanan, Richard Kaufmann, Bu-Sung Lee, Filippo Gioachin, Verdi March, and Chun Hui Suen

Transcript of HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress...

Page 1: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012

Dejan Milojicic

Hewlett-Packard Laboratories

[email protected]

Work with

Nigel Cook, Paolo Faraboschi, Abhishek Gupta, Laxmikant V. Kalé, Sudarsun Kanan,

Richard Kaufmann, Bu-Sung Lee, Filippo Gioachin, Verdi March, and Chun Hui Suen

Page 2: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 2

Who is Dejan Milojicic? • 27 Years of Experience in three world class industrial labs

−HP Labs, Palo Alto, CA,1998-now

−Open Software Foundation Research Institute, Cambridge, MA, ‘94-’98

−Research Institute “Mihajlo Pupin”, Belgrade, Serbia, 1983-1991

• Areas of expertise −Distributed systems; service management; cloud and distributed computing;

HPC; support automation; and systems software

• Education −PhD (93) Kaiserslautern Germany; MSc (86), BSc (83) Belgrade University

• Professional Activities −EIC IEEE Computing Now, IEEE Internet Computing Edit. Board, ICAC’12

general chair, IEEE Services Program Chair, Many PCs, etc.

−IEEE Computer Society, Board of Governors, 2014 President Nominee

−IEEE Fellow, ACM Distinguished Engineer, USENIX member

−Over 120 publications; 11 granted patents,19 filed

Page 3: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 3

HP LABS AROUND THE WORLD

PALO ALTO

BRISTOL

ST. PETERSBURG

HAIFA

BEIJING

BANGALORE

SINGAPORE

Page 4: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 4

Content Transformation

HP LABS RESEARCH PORTFOLIO

Cloud

Information Management

Digital Commercial Print

Sustainability

Immersive Interaction

Analytics

Intelligent Infrastructure

Page 5: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 5

What is High Performance Computing?

• Revenue around $10B

• Supercomputers; divisional, departmental, workgroup servers

• Market: servers, storage, middleware, applications, service

• Verticals

− Bio; Chemical; Lifesciences; Medical; Pharmaceutical;

− National Security and Homeland Defense; university and academic

− Automotive; Gas and Oil; Financial; Weather Forecasting

− CAD; CAE; Electronic Design and Analysis (EDA); Geo Engineering

− Gaming, Digital Content and Entertainment, etc., etc.

• Vendors: Bull, Cray, DDN, Fujitsu, HP, IBM, Intel, Mellanox, NEC, NVIDIA, Panasas, SGI, etc., etc.

• ISVs: Adaptive Computing, Altair, Platform Computing, etc., etc

Page 6: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 6

What is Cloud

• Revenue:

−Public Cloud: $21.5B (2010) $72.9B (2015) CAGR 27.6%

− IT infrastructure public Cloud $4.2B (2010) $10.3B (2015) CAGR 19.6%

− IT infrastructure private Clouds $4.8B (2010) $11B (2015) CAGR 17.8%

−Self-built Clouds less than $1B revenue in 2015.

• IaaS:

− Rackspace, IBM Cloud, Dell, HP, Hosteurope, LayeredTech, LongJump, NetApp, Newservers, 10gen, ReliaCloud, Symetriq, Skytap, Zoho,

• PaaS:

− Google AppEngine, Microsoft Azure, VMware CloudFoundry, IBM Workload deployer, CloudBees, SalesForce Heroku, RightScale Zend, Zoho creator,

Page 7: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 7

5 Key Megatrends

7

• Urbanization

• Increased Middle Class

• Aging Population

• Health Awareness

Demographics

• Brands

• Markets

• Competition

• Partners

• Suppliers

Globalization

• Client Interactions

• Referencing

• Advise

• Advertising

Social Media

• Product features

• Usage

• Communication

• Business Models

Mobility

Sustainability

Page 8: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 8

Accelerating Innovation & Change

8

The Internet

Client/Server

Mobile, Social,

Big Data & The Cloud

Mainframe

Database

ERP

CRM

SCM

HCM

HCM

PLM

MRM

Amazon Web Services

OpSource

IBM

GoGrid

Rackspace

Joyent

Hosting.com

Tata Communications

Datapipe

PPM

Alterian

Hyland LimeLight

NetDocuments

NetReach

OpenText

PaperHost

Xerox

Google

HP

Microsoft SLI Systems

EMC

IntraLinks

Jive Software

Qvidian

Sage

salesforce.com

SugarCRM

Volusion

Xactly

Zoho

Adobe

Avid

Corel

Microsoft

Paint.NET

Serif

Yahoo

CyberShift

Saba

Softscape

Sonar6

Ariba

Yahoo!

Quadrem

Elemica

Kinaxis

CCC

DCC

SCM

Cost Management

Order Entry

Product Configurator

Bills of Material Engineering

Claim Processing

Inventory

Manufacturing Projects

Quality Control

Business

Education

Entertainment

Games

Lifestyle

Music

Navigation

News

Photo & Video

Productivity

Reference

Social Networking

Sport

Travel

Utilities

every 60 seconds

400,710 ads requests

2000 lyrics played on Tunewiki

1,500 pings sent on PingMe

34,597 people are using Zinio

208,333 minutes Angry Birds played

23,148 apps downloaded

Unisys

Burroughs

Hitachi

NEC

Bull

Fijitsu

ADP VirtualEdge

Cornerstone onDemand

CyberShift

Workbrain

Kenexa Saba

Softscape

Sonar6

SuccessFactors

Taleo

Workday

Workscape

Exact Online

FinancialForce.com

Intacct NetSuite

SAP

NetSuite

Plex Systems

Cash Management

Accounts Receivable

Fixed Assets Costing

Billing

Time and Expense

Activity Management

Payroll

Training

Time & Attendance

Rostering Sales tracking &

Marketing

Commissions Service

Data Warehousing

98,000 tweets

Finance

box.net

Facebook

LinkedIn

TripIt

Pinterest

Zynga

Zynga

Baidu

Twitter

Twitter Yammer

Atlassian

Atlassian

MobilieIron

SmugMug

SmugMug

Atlassian

Amazon

Amazon

iHandy

PingMe

PingMe

Associatedcontent

Flickr

Snapfish

YouTube

Answers.com

Tumblr.

Urban

Scribd.

Pandora

MobileFrame.com

Mixi

CYworld

Qzone

Renren

Xing

Yandex

Yandex

Heroku

RightScale

New Relic

AppFog

Bromium

Splunk

CloudSigma

cloudability

kaggle

nebula

Parse

ScaleXtreme

SolidFire

Zillabyte

dotCloud

BeyondCore

Mozy

Viber

Fring Toggl

MailChimp

Quickbooks

Hootsuite

Foursquare

buzzd

Dragon Diction

eBay SuperCam

UPS Mobile

Fed Ex Mobile

Scanner Pro

DocuSign

HP ePrint

iSchedule

Khan Academy

BrainPOP

myHomework

Cookie Doodle

Ah! Fasion Girl

Page 9: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 9

Speed in Business is Increasing Dramatically

9

107 106 105 104 1,000 100 10 1 Seconds

Trading analytics

Airline operations

Call center inquiries

Track financial position

Supply chain updates

Mail/express/fax/e-mail

Document transfer

Phone activation

Refresh data warehouse

Trade settlement

Build-to-order PC

30 minutes 5 seconds

20 minutes 30 seconds

8 hours 10 seconds

1 day 10 seconds

1 day 5 minutes

30 seconds 3 days

3 days 45 seconds

1 month

1 day 5 days

6 weeks 24 hours

3 days 1 hour

1 hour

Server Provisioning 8 weeks 5 Minutes

Page 10: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 10

HPC in the Clouds

• Use of clouds for HPC growing, limited to small scale, test&dev

• Amazon built a top-500 supercomputer in its cloud

−7 k cores, 41.82 teraflops, 231st fastest supercomputer (at the time)

−with Linux on Intel Xeon X5570 with a 10 Gig Ethernet interconnect.

−de-provisioned soon after, demonstrated supercomputer at $1.60/node/hour

• At high-end HPC, US Department of Energy preparing Exascale program, and so are governments in Europe, China and Japan

−these are boundaries of high-end HPC, evolving as high-end data centers

−major differences slower interconnects, less powerful computation nodes

−similarity is in power, cooling, and packaging

Page 11: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 11 11

Characteristic Year

2010 2015 2018

Power 6MW 15MW 20MW

Nodes # 18,700 5,000 100,000

Node concurrency 12 ~1,000 ~10,000

Interconnect BW 1.5GB/s 1TB/s 2TB/s

MTTI Day ~Day ~Day

HPC Evolution towards Exascale

Page 12: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

FLAVORS OF HPC

• “ELITE SUPERCOMPUTING”

• Will never move to the cloud

• “MPC” (MEDIUM PERFORMANCE COMPUTING)

• Has already moved (or is quickly moving) to the cloud

• EVERYTHING IN BETWEEN AND THE “NEW HPC”

• It will all eventually move to the cloud

• But, the cloud will have to RADICALLY change before that happens

Cloud

Processor HPC optimized Commodity

Network

HPC specific IBM BlueWaters,

BlueGene, Cray

XMT

Cray XT5

Infiniband x86 + GPGPU +

QDR IB

x86 + QDR IB

Commodity N/A x86 + Gbit

Ethernet clusters

High end HPC

Page 13: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 13

Motivation: Why Clouds for HPC ?

• No startup/maintenance cost, cluster create time

• Time-to-solution - job wait time

• Elastic Resources, no risk e.g. in under-provisioning

• Power savings, prevents underutilization

• Benefits of virtualization

− Flexibility and Customization

− Security and Isolation

− Migration

− Resource Control

• Hence, a cost-effective and timely solution

− e.g. substitute/addition when Supercomputers are heavily loaded

Page 14: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 14

How is Cloud Different than HPC System?

• Interconnect: GigE, at best 10GigE vs Infiniband

• Virtualization vs tuned Operating System

• Cost vs grants and quotas

• VM instances (at best clusters) vs job submission system

• Software Architecture

• Security, trust, governance, ..….

• Reliability, availability, support, ……

Page 15: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 15

Bottlenecks in the Cloud: Micro-benchmarks

• Inferior network performance

• Virtualization overhead on network

• Noise (non-tuned OS, presence of hypervisors)

• Optimized for running business/web apps NOT HPC

• Slow Interconnect major bottleneck: 1-2 orders of magnitude

Page 16: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 16

What are the bottlenecks

NAMD CPU% for 60 cores on Eucalyptus Cloud, white portion is idle time

Page 17: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 17

HPC apps sensitivity (Noise/Virtualization)

Difference in core 0 and core 1 performance

– Network emulating driver process runs on core 0 of VM – consumes more CPU cycles as application communicates more

Page 18: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 18

Economics: Why HPC in the Cloud?

• Variable usage in time (resulting in lower utilization)

• Trading CAPEX for OPEX

• Shift towards a delivery model of Software as a Service

• Cloud users perspective

− Use cloud when their applications fit the profile above

• Cloud providers perspective

− Aggregated resource utilization of all tenants can sustain a profitable pricing model compared to substantial infrastructure investments required to offer a cloud

Page 19: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 19

Economics: Why not HPC in the Cloud?

• HPC systems typically highly utilized

− Queue-based approach keeps supercomputer busy 24x7

• Mismatch between HPC demands and cloud offerings

− HPC application performance sensitive to interconnect but typical deployment in cloud is commodity Ethernet (1Gbps moving to 10Gbps)

− Noise caused by virtualization and multi-tenancy can significantly affect HPC applications in terms of performance predictability and scalability.

• CAPEX/OPEX argument is less clear-cut for HPC users

• Software-as-a-Service offering rare in HPC to date

Page 20: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Cloud Challenge #3: The Cost Of Energy

MUST ADDRESS ENERGY ACROSS THE BOARD: COMPUTE, COMMUNICATION, STORAGE

Page 21: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

56X in seven years 16X

Moore’s law

“Physics of data, Myers, Google”

Cloud Challenge #2: Explosion Of Data

Data on higher exponential growth than compute

Online data: 5EB280EB in last 7 years (95%/yr) vs. 40% Moore

Page 22: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Cloud challenge #1: Commodity Interconnects

• HPC apps are latency bound

• Cloud commodity interconnects are inadequate

• Low-latency interconnects are not financially viable in the cloud

• Photonics to the rescue! Some of these issues may diminish as we move to optical interconnects

• The question remains whether we’ll see HPC specialized clouds…

Page 23: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 23

Outline

• Evaluating HPC in the Cloud

− Interconnects, virtualization, economics, power

• Use Case Genome Sequencing, Science as a Service

• Use Case in Financial Services

• Future trends

− Storage (NVRAM), photonics

• Summary

• What I will not be talking about

− Big data, GPGPU, Exascale, power, reliability

Page 24: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Interconnect

Abhishek Gupta and Dejan Milojicic, Proceedings of the 6th Open Cirrus Summit, Fall 2011, Atlanta, Georgia, best student paper award

Abhishek Gupta, Paolo Faraboschi, Bu-Sung Lee, Dejan Milojicic, Filippo Gioachin, Verdi March, and Chun Hui Suen, HPDC, to appear

Gupta, et al., “High Performance Computing Applications in the Cloud,” in submission.

Page 25: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 25

Past Research

• Focus on performance alone as the metric

• Mostly on small scale (up to few 10’s of cores)

• Focus on Amazon EC2

• MPI based

• Pessimistic results

Page 26: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 26

Research Goals/Questions

• What are the performance bottlenecks in the cloud?

• What application characteristics are crucial to identifying the suitable platform for an execution?

• Is it beneficial to run some applications on supercomputer and some on cloud rather than all on a single platform?

• What are economical aspects & use cases for HPC in cloud?

Page 27: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 27

Experimental Testbed

Virtualization Testbed

Page 28: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 28

Hypothesis

Run on Cloud

SC Cloud VM

Cluster

App’s Latency Sensitivity

GFLOPS/sec

Cost

$ vs Performance and Latency Sensitivity

Cloud suitable for some and not all HPC applications

• Application Characteristics

• Scale

• User preference – performance, cost

Run on SC

Page 29: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 29

Slowdown vs. Parallel Efficiency

0.5

1

2

4

8

16

32

0 0.2 0.4 0.6 0.8 1 1.2

Slow-down

Eucalyptus vs

Taub

Parallel Efficiency on Taub

Candidates for Cloud

0.5

1

2

4

8

16

32

0 0.5 1 1.5

Slo

w

do

wn

i… Parallel …

Efficiency E = S/P, where

P is the number of processors,

Speedup (S) is defined as: S = Ts/Tp where

Ts is the sequential execution time and

Tp is the parallel execution time.

Page 30: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 30

Benchmarks and Applications

• NAS Parallel Benchmarks class B (NPB3.3-MPI)

• NAMD - Highly scalable molecular dynamics

• ChaNGa - Cosmology, N-body

• Sweep3D - A particle in ASCI code

• Jacobi2D - 5-point stencil computation kernel

• Nqueens - Backtracking state space search

Page 31: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 31

Performance

31

Page 32: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 32

Performance Variability

• Coefficient of variation = Standard deviation/Mean

• Significant variability on cloud compared to supercomputers

• Variability increases as we scale up

Page 33: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Virtualization and Affinity

Abhishek Gupta, Paolo Faraboschi, Bu-Sung Lee, Dejan Milojicic, Filippo Gioachin, Verdi March, and Chun Hui Suen, HPDC, to appear

Page 34: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 34

Simplified Diagram of Virtualization Techniques

No virtualization

OS

Network Interface

OS Host OS

Network Interface

Network Interface

Host OS

Network Interface

Guest OS Guest OS

Container Plain VM Thin VM

User process

Page 35: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 35

Impact of Virtualization on App Performance

Ping-Pong NAMD ChaNGa

• Lightweight virtualization

−Thin VMs configured with PCI pass-through for I/O

−Containers (i.e., OS-level virtualization).

Page 36: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 36

CPU Affinity in Physical Infrastructure

Affinity No Affinity

User process

CPU

Page 37: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 37

CPU Affinity in Cloud

Application-level Affinity Full Affinity No Affinity

User process

vCPU

CPU

Page 38: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 38

Impact of CPU Affinity in Cloud Platform

Physical Machine (No CPU Affinity)

Thin VM (App-level Affinity) & Physical Machine (CPU Affinity)

All Configurations (CPU Affinity)

• Effect of CPU Affinity

− Application level (bind processes to the virtual CPUs)

− Hypervisor level (bind virtual CPUs to physical CPUs).

Page 39: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 39

App Performance with various CPU Affinities Thin VM: Impact of CPU Affinity on NAMD Thin VM: Impact of CPU Affinity on ChaNGa

Plain VM: Impact of CPU Affinity on NAMD Plain VM: Impact of CPU Affinity on ChaNGa

Page 40: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Economics

Avetisyan, A., Campbell, R., Gupta, I., Heath, M., Ko, S., Ganger, G., Kozuch, M., O’Hallaron, D., Kunze, M., Kwan, T., Lai, K., Lyons, M., Milojicic, D., Lee, H.Y., Soh, Ming., N.K., Luke, J.Y., Namgong, H., “Open Cirrus A Global Cloud Computing Testbed,” IEEE Computer, vol 43, no 4, pp 42-50, April 2010.

Campbell, R., Gupta, I., Heath, M., Ko, S., Kozuch, M., Kunze, M., Kwan, T., Lai, K., Yan Lee, H., Lyons, M., Milojicic, D., O’Hallaron, D., and Chai Soh Y., “Open CirrusTM Cloud Computing Testbed: Federated Data Centers for Open Source Systems and Services Research,” Proceedings of the USENIX HotCloud’09.

Page 41: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 41

Single site Cloud: to Outsource or Own?

• Medium-sized organization: wishes to run a service for M months

− Service requires 128 servers (1024 cores) and 524 TB

− Same as UIUC cloud site

• Outsource (e.g., via AWS): monthly cost

− Storage ~ $62 K

− Total ~ $136 K (using 0.45:0.0.4:0.15 split for hardware:power:network)

• Own: monthly cost

− Storage ~ $349 K / M

− $ 1555 K / M + 7.5 K (includes 1 sysadmin / 100 nodes)

• Breakeven analysis: more preferable to own if:

− M > 5.55 months (storage)

• Not surprising: Cloud providers benefit monetarily most from storage

− M > 12 months (overall)

• With underutilization of x%, still more preferable to own if:

− x > 33.3%

− Even with CPU util of 20%, storage > 47% makes owning preferable

41

Page 42: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 42

Economics: Why Cloud for HPC? (Revisited)

• HPC users in small-medium enterprises much more sensitive to the CAPEX/OPEX argument

− Startups with HPC requirements (e.g., simulation, modeling)

− Small medium enterprises with growing business and existing HPC infrastructure

• Ability to take advantage of a large variety of different architectures

− Better utilization at global scale

− Potential savings for consumers running HPC applications on most economical architecture while meeting performance expectations

Page 43: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 43

HPC Economics in the Cloud

Page 44: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 44

Cloud Bursting and Mapping

• Applications behave differently on different platforms,

− interesting cross-over points when considering cost

• Mapping Applications to platforms

− which application to burst

− which cloud to burst to

• Benefits

− Performance and cost

− Increased resource utilization

− Match between user expectations and application execution

Page 45: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 45

Related work • Studies on HPC in cloud

− Walker, He et al., Ekanayake et al.

• US DoE’s Magellan project

− compared conventional HPC platforms to Amazon EC2 and used real applications

− concluded that interconnect and I/O performance on commercial cloud severely limits performance and causes significant variability in performance across different executions.

− more cost-effective to run DOE applications on in-house supercomputers rather than on current public cloud offerings.

• Cost Evaluation for HPC in Cloud

− Napper and Bientinesi, Gupta and Milojicic

Page 46: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 46

Lessons and Conclusions

• A hybrid cloud-supercomputer platform environment can outperform its individual constituents

• Lightweight virtualization is important to remove overheads for HPC in cloud

• Application characterization in the HPC-cloud space is challenging but the benefits are substantial

Page 47: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Nigel Cook, Dejan Milojicic, Richard Kaufmann, and Joel Sevinsky

Proc. 7th Open Cirrus Summit, Beijing, China, 2012, best student paper award

n3phele: Open Science-as-a-Service Workbench for Cloud-based Scientific Computing

Page 48: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 48

100 GB/ run

21st Century BioScience Challenges

“To define the normal human microbiome, HMP researchers sampled 242 healthy U.S. volunteers, collecting tissues from 15 body sites in men and 18 body sites in women, [and processing] through all 3.5 terabases of genomic data”

Human Microbiome Project press release June 14, 2012 Processing

pipeline

innovation

Page 49: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 49

Existing Approaches

• Packaged Software builds

− BioLinux, QIIME

• Integrated Software & Workflow platforms

− CLOVR, GALAXY, YABI

• Software as a Service

− CLOTU, GALAXY, YABI, NCBI

• Research Clouds

− Magellan, DIAG

Main (Galaxy hosted )

Local Galaxy Install

Cloud

Your data sets are moderately sized Yes Yes Yes

Your computational requirements are moderate

Yes Yes Yes

You want to share your Galaxy objects with others

Yes Yes Yes

All needed Tools are installed on Main.

Yes ? Yes

Your data sets are very large No ? Yes

Your computational requirements are very large

No ? Yes

You have absolute data security requirements

No Yes Yes

Choice of Approach

from: Galaxy Web Site http://wiki.g2.bx.psu.edu/Big%20Picture/Choices

Page 50: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 50

n3phele (nef-el’) approach

• All information content and computing on cloud

• Access via browser, nothing to install on laptop or tablet

• Existing software runs without being re-written

• No progr. language/OS constraints; No platform integration

• Independently develop, process based on cloud publishing

Decouple UI, orchestration from software publication, execution

Cloud paradigm not virtualization

IaaS Compute Clouds

meta-data

n3phele

requests

Directed execution & data transfer

commands history activity

orchestration

users

Independently published software

Cloud repository

Page 51: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 51

Cloud repositories

https:// n3phele .appspot .com

Execution Clouds/grids

S3

HP Cloud

/Open CIrrus

ec2Factory Open Stack Factory

qiime

agent

ami-24858

qiime

ami-1234

File copy

Requests xfer, execute, status

vm create/delete/monitor

agent

Amazon

EC2

commands

accounts

activities repositories

Java servlet instance

Google object store (replicated)

Instance VMs

Google App Engine Components

queue

Swift

Page 52: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 52

Demo

52

Page 53: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 53

Command Definition & Primitives

• JSON definition published by any user

• Defines command parameters, required I/O files, and sequence of actions

• Actions execute based on availability of their dependencies: values from other actions, and input files

• UI and orchestration automatically derived from JSON definition

• Commands execute inside a container on VM; n3phele ensures all command dependencies are available before execution commences

Actions

• Create n VMs

• Execute shell on VM

• Fork n copies of series of action

• Wait for all fork components to finish

• Copy a file to/from Repo and VM

Command: “qiime”

Inputs: “map”, “fasta” Outputs: “otu_table”

Params: “limit” default=0.9

Cloud: “ec2” Input: “map” is “~ubuntu/wf_da/map.txt” Actions:

createVM name=x, ami-24858 runShell name=init, agentURI=“$<x>.ip”, cmd=“cd wf_da; process.py –l $<limit> map.txt ”

Page 54: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 54

Microbial Study of Gas Field • Using QIIME open source toolset

published by CU Boulder Knight lab

• Complete analysis <$200

• Multi-core data set Roche 454 denoising

• QIIME core analysis of community

Orchestration Class (# instances/run)

Run (ms)

1 2 3 4 5 Average

Activity(1) 76.8 64.6 62.8 73.0 63.9 68.2

createVM(1) 11.8 3.90 3.9 3.6 3.58 5.4

fileXfer (7) 14.3 19.1 10.3 11.7 26.7 16.4

Iterator(1) 4.2 3.5 3.5 3.4 9.2 4.7

Join(1) .6 .4 .6 .4 .5 .5

shellExecution(8) 19.3 28.7 11.8 21.8 10.1 18.3

Total (s) 127.0 120.2 92.9 114.0 114.0 113.6

N3phele orchestration & delegated worker lifecycle management CPU Consumption (ms) for Small (4 Amazon instance/30 minute) Denoising Run (600 Mhz GAE CPU)

Page 55: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 55

Conclusion

• Demonstrated cloud-based workbench Architecture, Design and Implementation with separation of application publishing, workload execution & orchestration

• Cloud computing is cost effective for scientific analysis, providing on-demand access from anywhere

• Google app engine provides a good prototype, but more control over delivered performance may be required

• Acknowledgements: R. Knight, G. Caporaso, A. Gonzalez, P. Marshall, H. Tufo CU Boulder

• Cloud resources from: F. Meyer (ANL), Amazon, HP

• Work funded in part by NIH HG004872

Page 56: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Financial Vertical

“Cloud Computing, Implications on the Financial Industry,”

Presentation at the CIAB, a Financial Services Congress in Brazil, June 2010

Page 57: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 57

Financial Applications

• Key applications going forward − Risk management (driven by the demand to remedy past failures)

− Trading systems (driven by fluctuating resource demand)

• Key benefits − Flexibility of dynamic resource allocation (eg Monte Carlo simulation)

− Enables some of the smaller financial firms (eg hedge funds)

• State of adoption − Evolution of Grids already deployed by many banks

− Perceived security risks, but experiments underway (ML, BA)

− Grid software vendors also moving to Cloud (Platform, Data Synapse)

− North and Latin America leading in IT spending

Page 58: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 58

Risks and Opportunities in Financial Clouds

• Security (confidentiality, integrity, availability)

− No access to physical resources, multitenancy, different security models

− Data in Cloud reduces exposure, homogeneity simplifies auditing, automated security management, etc.

• Regulatory compliance

− Export rules, privacy rules, global coverage (across region)

− Automation already in place (single control point), awareness raised

• Ilities: performance, availability, business continuity

− Lack of QoS, SLA standards/enforcement, will providers go away

− Marketplace of service providers, competition, geographical distribution

• Data lock-in

− Network performance lacking, proprietary I/F, semantic models

− Promising optical networking research, standardization in progress

Page 59: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 59

Financial Services Predictions and Cloud Implications

• Banks will focus on delivering information to user

− Analytics, data warehousing in the Cloud for processing & storage

• Risk compliance and regulations will be important

− Huge consumer of processing

• New business models seeking profitability

− Long tail of services: prepaid cards for teens; interbank protocols; unstructured data community; mobile banking platforms, etc.

• Lower tier capital firms will rise

− Startup equivalent, ideal for the use of Cloud

• Driving efficiency out of IT

− Ultimate consolidation through the use of Cloud

Page 60: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

HPC in the Cloud Research and Testbeds

Page 61: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 61

Open Cirrus™ Cloud Computing Testbed Shared: research, applications, infrastructure (20K cores), data sets

Global services: sign on, monitoring, store. Open source stack (prs, tashi, hadoop)

Sponsored by HP, Intel, and Yahoo! (with additional support from NSF)

• 15 sites currently, target of around 20 in the next two years.

61

GaTech

China Telecom CESGA

Univ.

Univ. Ind. Research

University

China Mobile

HP Confidential

Chinese Academy of Sciences

Ind. Research

Ind. Research

Page 62: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 62

What kinds of research projects are Open Cirrus sites looking for?

• Open CirrusTM is seeking research in the following areas (different centers will weight these differently) •Datacenter federation

•Datacenter management

•Web services

•Data-intensive applications and systems

•Hadoop map-reduce applications

• The following kinds of projects are of less interest •Traditional HPC application development

•Production applications that just need lots of cycles

•Closed source system development

62

Page 63: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 63

Cloud Sustainability Dashboard (CSD)

Open Cirrus

Site

Economical ($) Ecological Social

IT cooling ntwk support econo.

overall

CO2

(tonnes-eq)

water

(mill. Gal)

Resource

Use (GJ-eq)

ecolog.

overall

State of

devt.

Risk of

instability

social

overall

Site 1 $0.72 $0.35 $0.16 $0.43 6.0 2.6 83 High Low

Site 2 $1.27 $0.59 $0.21 $1.11 6.8 3.3 96 High Very Low

Site 3 $1.05 $0.47 $0.12 $1.07 5.9 2.3 81 High Low

Site 4 $0.75 $0.35 $0.12 $0.61 6.1 2.7 85 High Very Low

Site 5 $0.27 $0.13 $0.05 $0.09 4.3 2.4 59 Low High

Site 6 $1.82 $0.77 $0.11 $1.17 10.2 4.3 142 High Low

Site 7 $1.23 $0.54 $0.11 $0.98 15.0 4.4 192 High Low

Site 8 $0.55 $0.26 $0.10 $0.16 6.9 2.6 95 Med. Low

Site 9 $1.01 $0.44 $0.10 $0.83 5.3 2.5 74 High Very Low

Bricks-and-

Mortar (US) $0.58 $0.70 $0.12 $0.83 9.0 2.1 127 High Very Low

Page 64: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 64

CSD Summary

•A systematic approach for representing and assessing sustainability of Clouds

• Derived from a comprehensive model (economical, ecological, social)

• Automated, real-time Cloud Sustainability Dashboard

• Express, assess and display run-time sustainability of Cloud & Cloud services

• Preference-based customization

• Opportunities for integration with different enterprise tools

EvaluationSustainability Models

The CloudHP

Yahoo

UIUC

Intel

KIT

IDA

MIMOS

RAS

ETRI

internet

Compute, Network, Storage, Application, Power, Cooling

Data Center or Cloud

Visualization

Dashboard

Sustainability-aware

Management

Sustainability Metrics(Energy, Cost, CO2, Water, Risk…)

Page 65: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 65

Workshop on Cloud Services, Federation, and the 8th Open Cirrus Summit In conjunction with “ICAC’2012, San Jose, 18-20 Sep. 2012"

− http://fedcloud.cyberaide.org/, contact: [email protected]

Deadlines:

− Submission (Jul 14, ‘12); Notification (Jul 31, ‘12); Final (Aug 14, ’12)

Summary: This workshop will build upon the success of the prior Open Cirrus events and the prior Open Cloud Consortium events. The goal is to help building a community for those responsible for operating clouds and cloud testbeds, as well as those interested in designing new cloud services Co-sponsors: Open Cirrus, Open Cloud Consortium, and FutureGrid

Goals:

− Bring together researchers and practitioners to discuss the newest ideas and challenges in cloud services and federated cloud computing

− bring together those responsible for designing, managing, and operating clouds services so that they can share experiences with each other

− The workshop also welcomes users with requirements for new cloud services

− We are particularly interested in cloud services for federating clouds

Page 66: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 66

Topics of interest

• Experiences, best practices, lessons learned from operating cloud services

• Testbeds for designing new cloud services

• Cloud services for federating clouds

• Management and provisioning of cloud services

• Health and status monitoring of cloud services

• Security of cloud services

• Requirements for new cloud services

• Reliability and fault tolerance of cloud services

• Cloud services that span public and private clouds

• Cloud services design: Intercloud; Federation; Identity; Cloud bursting; etc.

• Cloud services for emerging applications

• Applications utilizing such services

Page 67: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 67

Organizers

General Chair: Michael Kozuch (Intel, Open Cirrus)

Program Chairs: von Laszewski, Gregor (Indiana U, FG) Grossman, Robert (U Chicago, OCC)

Steering Committee

Grossman, Robert (U. Chicago, OCC) Keahey, Kate (ANL, Nimbus) Kozuch, Michael (Intel, Open Cirrus) Milojicic, Dejan (HP Labs) von Laszewski, Gregor (Indiana U, FG)

Program Committee

Brandic, Ivona (TU Vienna) Desai, Narayan (ANL, Magellan) Desprez, Frédéric (INRIA, Grid5000) Diaz, Javier (Indiana U, FutureGrid) Fitzgerald, Steve (Eucalyptus) Fox, Geoffrey (Indiana U, FutureGrid) Gavrilovska, Ada (GaTech) Grossman, Robert (U. Chicago, OCC) Keahey, Kate (ANL, Nimbus) Kozuch, Michael (Intel, Open Cirrus) Llorente, Ignacio M. (OpenNebula) McGeer, Rick (HP Labs) Milojicic, Dejan (HP Labs) Riedel, Morris (FZ Juelich, EMI) Toews, Everett (Cybera) von Laszewski, Gregor (Indiana U, FG)

Page 68: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Some of the Trends

“Memristos and Photonics,” Norm Jouppi’s presentation at ISC

Page 69: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

NEW Non-Volatile Memories: Memristors

Page 70: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 70

The Prediction of a New Circuit Element: The Memristor

L. O. Chua, IEEE Trans. Circuit Theory 18, 507 (1971)

RESISTOR

dv = R di

CAPACITOR

dq = C dv

INDUCTOR

dφ = L di

MEMRISTOR

dφ = M dq

dq /dt = i

/dt =

v

i q

v

φ

1827

1831

1745

1971 )](,[

)(tiwf

dt

tdw

)()](,[)( titiwRtv rigorous

definition

Ohm

Faraday Chua

Von Kleist

Quasi-static conduction eq.-

R depends on state variable w

Dynamical equation –

Evolution of state in time

Page 71: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Memristors

• Class of Resistive RAM (RRAM)

• Crosspoint memory allows 4F2 cell

• 2.8nm junctions demonstrated

• 36X denser than today

• Can fab with multiple layers

• 4 layers -> 1F2 cell

• Die can be stacked too Xbar layer

wiring layer

CMOS layer

Page 72: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

New Main Memory Technologies Don’t Come Along Every Day

• Core memory (1960)

• DRAM (1973)

Page 73: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Memory Hierarchies with NVMs

On-chip memory

(SRAM)

Off-chip memory

(DRAM)

Secondary Storage

(HDD)

Solid State Disk

(Flash Memory)

How to leverage NVMs in memory hierarchies?

Page 74: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

New NVRAMs Are Disruptive Technologies

• We could just use NVRAM to replace disks & main memory

• Keep software the same

• But

• Storage I/O can speed up by 1000X and become byte addressable

• Memory can be huge

• Memory can be nonvolatile

Don’t let a disruptive

technology go to waste!

Page 75: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Improving HPC I/O in Cloud with NVRAM

Using active NVRAM for I/O staging. Sudarsun Kannan, Ada Gavrilovska, Karsten Schwan, Dejan Milojicic, and Vanish Talwar, In Proceedings of the 2nd international workshop on Petascal data analytics: challenges and opportunities (PDAC '11).

Using Active NVRAM for Cloud I/O, Sudarsun Kannan, Dejan Milojicic, Ada Gavrilovska, Karsten Schwan, Hasan Abbasi and Vanish Talwar (GT/HP Labs/ORNL), 6th Open Cirrus Summit, 2011

Page 76: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 76

• Need for efficient HPC I/O in cloud: • HPC applications demand high I/O throughput

• I/O needs for checkpoint, output, and diagnostic data

• Low network bandwidth in cloud compared to supercomputers

• Substantial I/O data movement cost across network

• I/O storage in traditional HPC is distributed (e.g., Lustre)

• Further, for efficient use of I/O, data needs to be post processed, for example, application diagnosis, visualization, data compression

• State of the Art: Data Staging • Main focus to reduce impact due to slow disks

• Intermediate I/O Staging nodes and in situ data processing: I/O Staging - ‘a partition’ of nodes service I/O needs of large number of compute nodes

• Other ideas: replacing disks with SSDs

• Problems with above approach in Cloud? • Network data movement to staging nodes still a problem

• Huge memory requirements from intermediate staging nodes

Motivation

Page 77: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 77

Network

Interconnect

HDD HDD

File System (e.g. Lustre)

Compute

VMs C1 C3

SSD

Node local NVRAM

Node local NVRAM

C2

Approach : Node local NVRAMs

• Node local NVRAM for app I/O

• NVRAMs 100X faster than SSDs

• NVRAM used like an additional heap

• Fast I/O to node local NVRAM follow by async I/O to dedicated storage

• VM unused compute cycles used for I/O data post processing before network transfer (e.g., compression and data reorganization)

Page 78: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 78

0

500

1000

1500

2000

20 22 30 35

Tim

e (

Sec

)

I/O Data (GB)

Post processing - NVRAM(Sec)

Post processing- Staging(Sec)

For smaller data sizes, I/O

staging performs better

For larger data sizes,

active NVRAM performs

better

Analysis and Initial Results

• High I/O performance gains compared to DataStaging

• DataStaging – I/O throughput -1.2 GB/sec

• Node local NVRAM approach – 8 – 10 GB/sec (predicted)

• I/O + data post processing performance compared to DataStaging

• I/O data compression used as an case study for GTC application

• Initial results show 60 % - 80% app performance improvement using node-local NVRAM for I/O and post processing approach

Page 79: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Photonics

Page 80: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

80

Photonics Portfolio and Time Scales

meters

centimeters

millimeters

Now 1 Year 5 Years 10 Years 3 Years 7 Years

11 October 30, 2007 HP DARPA Exascale Study — HP Proprietary & Confidential

2017 Design: 256 cores

L1-I

Core 0

Core 1

Core 2

Core 3

L1-D

L1-D

L1-I

L1-I

L1-D

L1-D

L1-I

L1

↔ L

2 In

terf

ace

Arbitration

Data

&

Control

Broadcast

Off-Chip

ModulatorsModulators

ModulatorsModulators

Modulators

Modulators

ModulatorsModulators

Modulators

Detectors Detectors

0

123

N-1

N

N+1

6162

63

2 fixed drop

filters

64 variable

drop filters/

detectors

4-waveguide

bundles

Modulators Detectors

Splitters

Modulators Detectors

Splitters

Hub

MC

NI

Through Silicon Via Array

Direct

ory

L1

↔ L

2 In

terf

ace

My X

-ba

r

Co

nn

ectio

n

Pe

er

X-b

ar

Co

nn

ectio

n

L2

Cache

Laser

Star coupler 1 Star coupler 2

22 October 30, 2007 HP DARPA Exascale Study — HP Proprietary & Confidential

Optical DIMM Architecture

Processor Chip

2nd

Channel1st Channel

Optical

DIMMs

DRAM DRAM DRAM DRAM

OMB

DRAM electrical

signaling

Optical Waveguide/Fiber

Optical

DIMM

Splitter Modulator

Photo-

detector

Opto-

electronic

signal

conversion

DRAM electrical signaling

Optical

OMB

Page 81: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Near Term: Optical Multidrop Bus

• Replace electrical transmission line with optical waveguides

• Replace electrical stubs with optical taps

• Two Unidirectional buses: 12 bit wide @ 10Gb/s = 30GB/s

• Master broadcasts to each module on the bus

• Distribute optical power equally among modules

• Each module sends data back to master at full bus bandwidth

• Lower latency with reduced power

Master Module A Module B Module C Module D

M

Tx Rx Rx Tx Rx Tx Rx Tx Rx Tx

BS1

BS1

M BS2

BS2

BS3

BS3 BS4

BS4 12

12

Page 82: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

1x8 Fanout

1 2 3 4 5 6 7 8 M

3cm

Light beams from taps.

bus Last tap exits bus end IR camera image

VCSEL driven from BERT thru bias-tee

Light input

30 cm

Page 83: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 83

Injection molded 12 channel Hollow Metalized Waveguides

0.06dB/cm

12 channels

Coated waveguide

Page 84: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Optical Backplane Assembly

Tap Detail

window

guide pins

Tx input (12 ch)

Tap 1 Tap 2 Tap 3 Tap 4 Tap 5 Tap 6

12 waveguide output

Near field of HMWG

Waveguides

Page 85: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Eye diagrams from all 6 taps @ 10Gb/s

Tap 1

Tap 2

Tap 4

Tap 5

Tap 6

Tap 3

ch 1 ch 2 ch 3 ch 4 ch 5 ch 6 ch 7 ch 8 ch 9 ch 10 ch 11 ch 12

All 12 channels active

(after receiver) Optical eye @ 10Gbps

ER > 5dB

20 ps/div

PRBS 27 -1

BER < 10-15

Page 86: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Potential Applications of Optical Busses

• Network switch backplane

• Demonstrated at Interop ‘11

Page 87: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Long Term: Integrated Photonics

• The 2000 telecom bubble based on discrete optics

− Think pre-Noyce/Kilby era in electronics

− Components are measured in mm

− Hand alignment

− Expensive and not scalable

• Recent research is on integrated photonics

− Think post-Noyce/Kilby era in electronics

− Components are measured in a few mm

− Manufacture many thousands per die

− Advances in lithography -> better devices

Source: Newport Corp.,

Assembly Magazine,

September 2001

Page 88: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Ring Resonators

• A modulator – move in and out of resonance to modulate light on adjacent waveguide

• A switch – transfers light between waveguides only when the resonator is tuned

• A wavelength specific detector - add a doped junction to perform the receive function

One basic structure, 3 applications

SiGe Doped

Page 89: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Long Term: The Corona Manifesto

• Take full advantage of nanophotonics

• Don’t just replace today’s wires with optics

• Redesign from the ground up

• No off-chip or cross-chip electrical wires

• Restore balance: memory bandwidth scales

• All memory readily reachable from all cores

OCM

OCM

OCM

OCM

OCM

OCM

OCM

OCM

OCM

Corona compute socket

Cluster

0

Cluster

1

Cluster

63

Optical Crossbar

Fiber I/O’s to

OCMs or

Network

Package

Memory Controller/Directory/L2 Die

Processor/L1 Die

Analog Electronics Die

Optical DiepgcTSVs

Face to

Face Bonds

Heat Sink

Laser

pgcTSVs

sTSVs

Page 90: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Performance (LMesh/ECM = 1)

Applications that don’t fit in cache show 4-6X improvements with Xbar

Page 91: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

On-chip Network Power

Optics can reduce network power of aps that don’t fit in cache by 6X

Page 92: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

Optics Summary

• Optical Bus

• Can build today

• Distance not an issue

• Provides good fan-in and fan-out (>8)

• Can combine to form crossbars

• Integrated Photonics

• Has great long term potential

• Bandwidth scales to 1,000 threads

• Coherent shared memory still possible

• Low and uniform memory latencies

Fiber I/O’s to

OCMs or

Network

Package

Memory Controller/Directory/L2 Die

Processor/L1 Die

Analog Electronics Die

Optical DiepgcTSVs

Face to

Face Bonds

Heat Sink

Laser

pgcTSVs

sTSVs

Page 93: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 93

Outline

• Evaluating HPC in the Cloud

− Interconnects, virtualization, economics, power

• Use Case Genome Sequencing, Science as a Service

• Use Case in Financial Services

• Future trends

−Storage (NVRAM), photonics

• Summary

Page 94: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 94

Summary

• HPC On the Cloud

− Clouds increasingly used for HPC, but trail high end supercomputers

− Challenges: Interconnects, power, scale of data

− New technologies offer promise: photonics, NVRAM

• HP Labs research in support of HPC in Clouds

− Basic architectural work: memristor, photonics

− Understanding performance, scale, power

− Moving up the stack: Science as a Service

− Open Cirrus Cloud computing testbed and related research

Page 95: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice.

Thank you

Page 96: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 96 96

Questions…?

Page 97: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 97

Mapping Approach and cost benefits

Exploring the Performance and Mapping of HPC Applications to Platforms in the Cloud by Abhishek Gupta et al. in HPDC 2012

Page 98: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 98

Benefits

• Intelligent decisions to determine mapping between applications and platforms

• Prevention of overloading of one infrastructure while others may be less loaded

− Better utilization

− Proper match between demand and supply

• Match between user expectation and application execution

• Reduced wait time for an application

− Incoming mix of applications scheduled to members in the set of platforms rather than getting concentrated on one.

− Increased overall throughput

Page 99: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 99

Contributions

• We analyzed performance of HPC applications on cloud and compare it with a range of dedicated platforms

• We analyzes impact of virtualization on HPC apps and proposed techniques, such as thin hypervisors and OS-level containers, to mitigate performance overhead and noise (jitter)

• We discussed a few concrete scenarios of cloud deployments and show that small/medium-scale HPC users are the most likely candidates to benefit from an HPC cloud

Page 100: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 100

The HP Cloud Architecture

100

HP Portal Experience Catalog, Service Request Management, Account Management, Aggregation, Performance, Cost, Billing, Service Reporting

HP Private Cloud Licensed Products HP Managed Cloud Services HP Cloud Services

HP Converged Management & Security Policy, Orchestration, Monitoring & Security

HP OpenStack

HP Converged IaaS Controller

Hybrid Cloud Solutions Hybrid Dev/Test, IaaS, Information as a Service

Hybrid Cloud Solutions

Standalone Cloud Products

& Services

Common Architectural

Elements

Differentiated & 3rd Party Infrastructure

HP Managed Cloud

HP CIaaS (converged Infrastructure as a service)

Converged Infrastructure

Servers Network Storage

HP Public Cloud

HP CIaaS (converged Infrastructure as a service)

Converged Infrastructure

Servers Network Storage

HP Private Cloud

HP CIaaS (converged Infrastructure as a service)

Converged Infrastructure

Servers Network Storage

Traditional

Traditional Heterogeneous

Hardware & Software

Servers Network Storage

3rd Party Clouds

Autonomy Protect & Promote

Management

Data Services

IaaS …… Dev/

Test

PaaS CRM

IaaS Mgmt Dev/

Test

Analytics Email, Collab, Unified Comms

Security Private Dev/Test

Private IaaS Private Cloud w/ Application

Lifecycle Mgmt Autonomy Protect & Promote

3rd Party

HP Converged Information Idol10 - manage data & metadata, from archive targets

Page 101: HPC in the Cloud - Hewlett Packard Enterprise · HPC in the Cloud Keynote, IEEE 8th World Congress on Services Honolulu, Hawaii, June 26, 2012 Dejan Milojicic Hewlett-Packard Laboratories

© Copyright 2012 Hewlett-Packard Development Company, L.P. The information contained herein is subject to change without notice. 101

Federation Economics

• Federation can help contain demand overflow within itself

• Cost of outsourcing overflow to public Cloud is higher than to federation of 6 sites

• Cost reduces with size of federation increasing to 50

101

10

100

1000

10000

100000

100% 120% 150% 200% 350% 600% 1200% 2500% 5000% 10000%

Mo

nth

ly C

ost

in

$K

Utilization

Existing DC

Open Cirrus 6

Open Cirrus 50