Slides from the AWS tutorial at Supercomputing 2011. Related tutorial materials at: cloudsupercomputing.net

1. High
Performance
Cloud
Computing
Supercomputing 2011

3. Hello

5. Thank you

6. HPC
with
AWS

7. Understand the
services, tools and
patterns for
building
high performance
systems in the

8. AGENDA
SC11 - Monday 14th November, 2011
Cloud Concepts
Building Blocks
Technica l & Scientific Computing
L oosely Coupled Systems
Hands-on Session #1
Parallel Computation
Hands-on Session #2
Wrap up

10. There Will Be Code

11. Cloud
Concepts
A prelude

12. Consumer Seller
business business

14. Decades of experience
Operations, management and
scale

16. Programmatic access

18. Unexpected innovation

20. Blinding flash of
the obvious

23. Five years young

25. Infrastructure services

26. Compute Storage
Placeholder
Services
Databases
& Support

27. Idea Results

28. Idea Results
Heavy lifting

29. Scale
Redundancy Orchestrati
on
70%
Idea Results
Heavy lifting
Capacity
Management Procurement

30. 30%
Idea Results
Infrastructure

31. Idea Results
AWS

32. Idea Results
AWS

33. Five things I
wish I’d known
when I
was getting
started.

35. 1: Signing up

37. On the web

41. Free tier
For new customers:
aws.amazon.com/free

42. 750 hours of compute
10Gb network attached storage
5Gb object storage
750 hours of compute
Key/value store, notifications,
messaging

44. 2: Interacting

45. HTTP, REST, SOAP

46. API driven
HTTP, REST, SOAP

49. CLI

50. ec2-run-instances

51. ec2-terminate-instances

52. Java, Python, Ruby, .Net, PHP,
iOS and Android

53. SDK
Java, Python, Ruby, .Net, PHP,
iOS and Android

58. Management
console

61. Linux

63. Certificate-
based root
access

64. mza$ ssh -i web/us-east/aws-web.pem
root@ec2-204-236-247-169.compute-1.amazonaws.com
Last login: Wed Jun 22 11:15:20 2011 from 82.26.6.99
__| __|_ ) CentOS
_| ( / v5.4
___|___|___| HVMx64
Welcome to an EC2 Public Image
:-)
[root@ip-10-17-135-244 ~]#

66. Windows

68. Administrator
access

71. 3: Storage
options

72. Ephemeral
storage

73. Included
with
compute
Ephemeral
storage
Lost at
Not backed
terminatio
up
n

75. When it’s gone,
it’s gone

76. Hands-
on

77. Elastic Block Store
Hands-
on

78. Network Mount as
attached volume
Elastic Block Store
Snapshot Persistent

79. Hands-
on

80. S3
Hands-
on

81. Highly Highly
durable available
S3
Tolerant to
two
simultaneo

82. durability

83. 99.999999999%
durability

84. Objects in S3
Billions of objects 556B
600
450
300
150
0
Q4 2006 Q4 2007 Q4 2008 Q4 2009 Q4 2010 Q3 2011

86. 370,000 peak
transactions per second

88. Payment options

90. Pay as you go

92. Gb/month

94. ECU/hour

96. No minimum

98. No subscriptions

100. Pricing tiers

102. Consolidated
billing

103. Options

105. On-demand

107. Reserved
capacity

108. Hands-
on

109. Spot Market
Hands-
on

110. Bandwidth

112. Free inbound

114. Import/Export

116. Reduced
outbound

118. Pricing calculator

119. aws.amazon.com/calculator

121. 5. Availability
Zones

124. us-east-1
us-west-1
us-west-2
us-gov-west-1
eu-west-1
ap-southeast-1
ap-northeast-1

125. eu-west-1a
eu-west-1b eu-west-1c

126. Building
blocksservices
Foundational

127. Compute

128. Elastic Compute Cloud

129. EC2
Elastic Compute Cloud

130. Hands-
on

131. Elastic compute
infrastructure
Hands-
on

132. ECU:
Equivalent to 1.0 - 1.2 GHz 2007 Opteron
or 2007 Xeon

133. ECU:
EC2 Compute Unit
Equivalent to 1.0 - 1.2 GHz 2007 Opteron
or 2007 Xeon

135. Instance types

136. Clust
Micro
er
$0.02 $2.10

138. Standard (m1)
1 ECU. 1.7 Gb memory.
160 Gb ephemeral
storage.

140. High memory
(m2)
Up to 26 ECU. 8 cores. 68.4 Gb
memory.
1.69 Tb ephemeral storage.

142. High CPU (c1)
Up to 20 ECU. 8 cores. 7 Gb memory.
1.69 Tb ephemeral storage.

144. Higher
performance

146. MPI workloads

148. Bandwidth
intensive

149. Hands-
on

150. CC:
Cluster Compute
Hands-
on

152. 2 x Intel Xeon
5570
23 Gb memory
1.7 Tb disk
33.5 ECUs

154. HVM

155. 10 gig E

157. Placement groups

159. Full bisectional
bandwidth

161. Linpack

162. November 2010
Cores 7040
R max 41.82
R peak 82.51

163. November 2010
231

164. June 2011
451

165. November 2011

166. November 2011
42

167. WIEN2K Parallel
Performance
H size 56,000 (25GB)
Runtime (16x8 processors)
Local (Infiniband) 3h:48
Cloud (10Gbps) 1h:30 ($40)
1200 atom unit cell; SCALAPACK+MPI
diagonalization, matrix size 50k-100k
Credit: K. Jorissen, F. D. Villa, and J. J. Rehr (U. Washington)

169. GPU computation

170. Hands-
on

171. CG:
Cluster Compute
with gpGPU
Hands-
on

172. 2 x NVIDIA
M2050

173. 2 x Intel Xeon 5570
23 Gb memory
1.7 Tb disk
2 x NVIDIA
M2050

177. Flexible cluster
control

179. API

180. Hands-
on

181. SGE
Hands-
on

183. LSF

185. Condor

187. Rocks+

189. Slurm

190. Included with all instances
and block storage

191. CloudWatch
Included with all instances
and block storage

193. Custom metrics

194. Storage

195. Simple Storage Service

196. S3
Simple Storage Service

198. Files in directories

200. Objects in
buckets

201. http://s3.amazonaws.com/bucketname/objectid
http://bucketname.s3.amazonaws.com/objectid

202. https://s3.amazonaws.com/bucketname/objectid
https://bucketname.s3.amazonaws.com/objectid

203. 5Tb

204. Large object
support
5Tb

206. Parallel uploads

208. Import/Export

210. Managed
encryption

211. 99.99% durability

212. Reduced
redundancy
storage
99.99% durability

213. Elastic Block Store

214. EBS
Elastic Block Store

216. Flexible, off-
instance block
storage

217. 1Gb to 1Tb

218. Scalable
1Gb to 1Tb

219. Exposed as a device

220. Attached to a
running instance
Exposed as a device

222. Snapshot to S3

223. Hands-
on

224. Public Datasets
Hands-
on

225. Databases

227. Databases on EC2

229. Oracle and MySQL

230. Managed. High availability.
Read replicas.

231. Relational
Database Service
Managed. High availability.
Read replicas.

232. High scale. Highly available.
Key/attribute store

233. SimpleDB
High scale. Highly available.
Key/attribute store

235. No server to
provision or
manage.

237. Perfect for
metadata

238. Messaging &
notifications

239. Hands-
on

240. Simple Queue
Service
Hands-
on

241. Hands-
on

242. Simple
Notification
Service
Hands-
on

243. Technical &
Scientific
Computing

245. Elasticity

246. Research is
bursty

247. Traditional
capacity is
static

248. Capacity
Predicted capacity
Estimated
demand
Time

249. Capacity
Infrastructure
Infrastructure
Investment Estimated
demand
Barrier to
entry
Time

250. Capacity
Infrastructure
Real
demand
Time

251. Capacity
Elastic
capacity
Real
demand
Time

253. Rapid response

255. Removing
constraints

256. Research is
constrained

257. Constrained by
static
infrastructure

258. Unconstrained

259. Larger systems,
more molecules,
more stars,
higher order
species...

260. Unconstrained
by scale

261. 30,000 cores

262. Unconstrained
by time
Upcoming conference, grant
submissions, impatience,
exploratory “spike”

263. 1 core
for
100 hours

264. 100 cores
for
1 hour

265. 10k cores
in
45 minutes

266. Unconstrained
by cost

267. Optimising for
price

268. On-demand

269. Reserved
capacity

270. 1&year&term&
Usage Fee One-time Fee Total Savings
Option 1 $1493 - $1493 -
On-Demand only
Option 2 $1008 $227 $1234 ~20%
On-Demand +
Reserved
Option 3 $528 $455 $983 ~35%
All reserved
Total&Cost&for&1&Year.term&of&2&applica4on&servers,&steady&state&usage&

271. 3&years&term&
Usage Fee One-time Fee Total Savings
Option 1 $4479 - $4479 -
On-Demand only
Option 2 $3024 $350 $3374 ~30%
On-Demand +
Reserved
Option 3 $1584 $700 $2284 ~50%
All reserved
Total&Cost&for&3&Year.term&of&2&applica4on&servers,&steady&state&usage&

272. 450"
On#Demand# 1*year#RI# 3*year#RI#
400"
350"
300"
250"
200"
2
150"
100"
50" 1
0"
1" 2" 3" 4" 5" 6" 7" 8" 9" 10" 11" 12" 13" 14" 15" 16" 17" 18" 19" 20" 21" 22" 23" 24"
on-demand vs. reserved instances

274. Spot Instances
Hands-
on

276. Placeholder

277. On-demand +
Reserved +
Spot

278. “20th Century
architectures”

279. Driven by
analysis
metrics

281. Increasing
usability

282. Amazon
Machine Image

283. Community
AMIs

284. http://www.cloudbiolinux.com/

285. http://usegalaxy.org/cloud

286. Reproducibility

287. Detailed
logging
For S3 access

288. Application
logging
Archive to S3 for durability

289. Automation

290. Application tier
Code Configuration

291. Application tier
Code Configuration

292. Application tier
Code Configuration
Service tier
Integration
Operating system
settings
Services +
Launch configuration
configuration

293. Application tier
Code Configuration
Service tier
Integration
Operating system
settings
Services +
Launch configuration
configuration

294. Application tier
Code Configuration
Service tier
Integration
Operating system
settings
Services +
Launch configuration
configuration
Infrastructure tier
AMIs Architecture Multi-AZ
Scaling rules Security groups Middleware

295. Value baked
into each tier

296. Service tier
Configuration & optimization
Technology choices

297. Infrastructure
tier
Architecture. Configuration.

298. Automation
maximises
this value

299. CloudFormation
Hands-
on

300. Template

301. Template
Defines a full infrastructure stack

302. Auto-scaling
RDS
EC2 SNS
SimpleDB
EBS SQS
Resources
Elastic Beanstalk CloudWatch
Security groups Tags

303. Template CloudFormation
Provisioned
resources

304. Complete
definition
Atomic, idempotent provisioning.

305. JSON
Declarative language

306. {
"AWSTemplateFormatVersion" : "2010-09-09",
"Description" : "Create an EC2 instances",
"Parameters" : {
"KeyName" : {
"Description" : "Name of an existing EC2 KeyPair to enable SSH access to the instance",
"Type" : "String"
}
},
"Mappings" : {
"RegionMap" : {
"us-east-1" : {
"AMI" : "ami-76f0061f"
},
"us-west-1" : {
"AMI" : "ami-655a0a20"
},
"eu-west-1" : {
"AMI" : "ami-7fd4e10b"
},
"ap-southeast-1" : {
"AMI" : "ami-72621c20"
},
"ap-northeast-1" : {
"AMI" : "ami-8e08a38f"
}
}
},
"Resources" : {
"Ec2Instance" : {
"Type" : "AWS::EC2::Instance",
"Properties" : {
"KeyName" : { "Ref" : "KeyName" },
"ImageId" : { "Fn::FindInMap" : [ "RegionMap", { "Ref" : "AWS::Region" }, "AMI" ]},
"UserData" : { "Fn::Base64" : "80" }
}
}
},
"Outputs" : {
"InstanceId" : {
"Description" : "InstanceId of the newly created EC2 instance",
"Value" : { "Ref" : "Ec2Instance" }
},
"AZ" : {
"Description" : "Availability Zone of the newly created EC2 instance",
"Value" : { "Fn::GetAtt" : [ "Ec2Instance", "AvailabilityZone" ] }
},
"PublicIP" : {
"Description" : "Public IP address of the newly created EC2 instance",
"Value" : { "Fn::GetAtt" : [ "Ec2Instance", "PublicIp" ] }
}
}
}

307. {
"AWSTemplateFormatVersion" : "2010-09-09",
"Description" : "Create an EC2 instances", Headers
Parameters
"Parameters" : {
"KeyName" : {
"Description" : "Name of an existing EC2 KeyPair to enable SSH access to the instance",
"Type" : "String"
}
},
"Mappings" : {
"RegionMap" : {
"us-east-1" : {
"AMI" : "ami-76f0061f"
},
"us-west-1" : {
Mappings
"AMI" : "ami-655a0a20"
},
"eu-west-1" : {
"AMI" : "ami-7fd4e10b"
},
"ap-southeast-1" : {
"AMI" : "ami-72621c20"
},
"ap-northeast-1" : {
"AMI" : "ami-8e08a38f"
}
}
},
"Resources" : {
"Ec2Instance" : {
"Type" : "AWS::EC2::Instance",
Resources
"Properties" : {
"KeyName" : { "Ref" : "KeyName" },
"ImageId" : { "Fn::FindInMap" : [ "RegionMap", { "Ref" : "AWS::Region" }, "AMI" ]},
"UserData" : { "Fn::Base64" : "80" }
}
}
},
"Outputs" : {
"InstanceId" : {
"Description" : "InstanceId of the newly created EC2 instance",
"Value" : { "Ref" : "Ec2Instance" }
},
Outputs
"AZ" : {
"Description" : "Availability Zone of the newly created EC2 instance",
"Value" : { "Fn::GetAtt" : [ "Ec2Instance", "AvailabilityZone" ] }
},
"PublicIP" : {
"Description" : "Public IP address of the newly created EC2 instance",
"Value" : { "Fn::GetAtt" : [ "Ec2Instance", "PublicIp" ] }
}
}
}

308. Bootstrap
Hands-
on

309. Chef & Puppet

310. Hands-
on

311. Elastic MapReduce
Hands-
on

313. Hadoop for data
intensive analytics

315. Painful at scale

316. S3
Input data

317. S3
Input data
Code Elastic
MapReduce

318. S3
Input data
Code Elastic Name
MapReduce node

319. S3
Input data
Code Elastic Name
MapReduce node
Elastic
cluster

320. S3
Input data
Code Elastic Name
MapReduce node
HDFS
Elastic
cluster

321. S3
Input data
Code Elastic Name
MapReduce node
Queries
HDFS
+ BI
Via JDBC, Pig, Hive
Elastic
cluster

322. S3
Input data
Code Elastic Name Output
MapReduce node S3 + SimpleDB
Queries
HDFS
+ BI
Via JDBC, Pig, Hive
Elastic
cluster

323. S3
Input data
Elastic Output
MapReduce S3 + SimpleDB

324. Hands-
on

325. Spot
Hands-
on

327. Enabling
collaboration

328. Data

329. Lots of data

330. Lots of data,
lots of uses

331. Lots of data,
lots of uses,
lots of users

332. Lots of data,
lots of uses,
lots of users,
lots of locations

333. Force
multipliers

334. Maximise value
of data

335. Data close to
compute

336. Import/Export

337. Direct Connect

338. AMIs,
Snapshots,
CloudFormation
Hands-
on

339. Public Datasets
Hands-
on

341. Ensuring
security

342. Shared
responsibility

343. Requirement
based access

344. Certification

345. ISO 27001

346. SAS70 Type II

347. Service
Organisation
Controls (SOC
1)
SSAE 16 and ISAE 3702

348. FISMA Moderate

349. HIPAA

350. ITAR
AWS GovCloud (US)

351. Data access
control
Detailed logging

352. Data stays local

353. Identity &
Access Control
Hands-
on

354. Account

355. Account
DBA Developer Sys admin Finance
Roles

356. Account
DBA Developer Sys admin Finance
Roles
Sally
Robert
Users
Chris

357. Security credentials
Multifactor authentication
Management console access
Data read/write access
API level access

358. Account
DBA Developer Sys admin Finance
Roles
Sally
Robert
Users
Chris

359. Networking
controls

360. Virtual Private
Cloud

361. Virtual network
topology

362. IP address range
Public and private subnets
Routing tables
Network gateways

363. Network access
control

364. Inbound ACLs
Outbound ACLs
IPsec VPN

365. Public subnet
Public facing website

366. Public subnet
Network ACLs + security groups
Private subnet
Multi-tier applications

367. Public subnet
Private subnet
IPsec VPN
On-premise
Extend your data centre

368. Private subnet
IPsec VPN
On-premise
Extend your data centre

369. aws.amazon.com/security

370. End of Part One

371. cloudsupercomputing.net/
tutorial

372. aws.amazon.com/
awscredits

373. Part Two

374. Hands-on:
Loosely coupled
systems

375. High scale,
loosely coupled
system

376. Embarrassingly
parallel

377. Decoupled,
batch
workflows

378. Tasks
Instances

379. Tasks
Queue
Instances

380. Tasks
Queue
Instances

381. Tasks
Queue
Instances
Increase
instance
size

382. Tasks
Queue
Instances
Increase
instance
size

383. Tasks
Queue
Instances
Increase
instance
count

384. Tasks
Queue
Instances
Results
Store

385. Tasks
Queue
On-premise
Instances
Results
Store

386. Tasks
Queue
On-premise
Instances
Results
Store

387. Tasks
Queue
On-premise
Instances
Results
Store

388. Batch
processing
Monitoring. Auto-scaling. Queuing.
Spot. Automation.

396. Configure

401. 150

406. Autoscaling.
Automation.
Don’t forget to shut down your instances!

407. Hadoop with
Elastic
MapReduce
Native. Streaming interface. Hive.
Spot with EMR.

408. Advanced EMR
with Myrna
Bioinformatics tools and large
datasets.
Thanks to Ben Langmead.

409. $100

410. Credentials
Account -> Security Credentials
Access key, secret key, account
number

411. AWS staff

412. cloudsupercomputing.net/
tutorial

413. Hands-on:
Parallel
Computation

414. Tightly coupled
systems

415. 10 gig E

416. 64 core parallel
cluster
CC1. Custom AMI. EBS. Monitoring.
MIT StarCluster. CloudFormation.

423. Multi-GPU
CG1. CUDA 4. Compile & execute.
Benchmark against CPU.

424. OpenFOAM
Computational Fluid Dynamics
with CC1 on EC2.

425. cloudsupercomputing.net/
tutorial

426. AGENDA
SC11 - Monday 14th November, 2011
Cloud Concepts
Building Blocks
Technica l & Scientific Computing
L oosely coupled systems
Hands-on Session #1
Parallel computation
Hands-on Session #2
Wrap up, drinks

427. Understand the
services, tools and
patterns for
building
high performance
systems in the

428. YOU ARE CORDIALLY INVITED TO THE
Amazon Web Services
S C 11 B A S H
NETWORKING, DRINKS and GOODIES
BOOTH #6202

429. aws.amazon.com/about-aws/sc11

430. Thank you!

431. Questions & comments:
matthew@amazon.co
m
@mza on Twitter