Elastic HPC Extending the Cluster into the Cloud Ruth Lynch, Research IT Service 13 th November 2009.

Elastic HPCExtending the Cluster into the Cloud

Ruth Lynch, Research IT Service

13th November 2009

HPC in UCD

Research IT Service 2005

HPC is Core Service Phaeton Cluster -

Community & Shared Services

11% ICHEC Stokes Cluster 2008  

What's the problem?

Physical HPC Cluster is inelastic Users sometimes need more resources than are

available Adding resources can be a slow process Capital intensive process

Possible Solutions

"Pre-order" hardware  Buy compute time on other

facilities Investigate  "The Cloud"

Which Cloud?

Cloud options Locations Different definitions

"a style of computing where massively scalable IT enabled capabilities are delivered 'as a service' to external customers using Internet technologies" - Gartner Inc. 

Getting on the Cloud

Amazon Web Services No technical issues How to pay for the services?

Credit Card? Prepaid credit card? Prepay for Amazon credit? Get billed by Amazon? ....... 

“All Aboard” the Cloud

Contacted Amazon Applied for a research grant with Amazon Got a small grant

Started building...

Technical Specifications

Compute Nodes Up to 20 x Large Instances AMD Opteron 2218 HE 4 vCPU Cores 7GB of RAM CentOS 5.2 On Demand

Security

Users authenticate through Phaeton Login Node LDAP Authentication

 Encrypted Communications 128 bit AES encryption for the data link 160 bit SHA1 checksum for both the data and

control channels

Management

Cloud Nodes are managed by Phaeton Head Node MOAB from Adaptive Computing / Torque Specific queue for cloud based resources

Data Concerns

Data Protection Laws Data Transfer Rules Intellectual Property Laws Data Privacy

Data Concerns

Amazon gives choice of location EU-WEST-1A & EU-WEST-1B

Data not "stored" in the cloud Computation only in the cloud

UCD CampusAWS

Computational Results

Real users with real jobs Results dependent on job and user type

0 1 2 3 4 5 6 7 8 90

5

10

15

20

25

30

35

Average Time per Step

Amazon Cloud

Phaeton

Job Number

Tim

e (

se

c)

What else?

AMD vs INTEL Licensing  Tuning

Cost comparison Currently appears more expensive than local

resources Approx. €9,500 per month for AWS Costs have recently fallen

The Benefits

Immediately available resources "Once off" usage charge for resources Relatively quick to set up Easy to maintain Viable option for occasional large scale single

jobs

Future Plans

UCD “Cloud” Incorporating existing local resources Access to multiple public cloud resources Seamless service provision Automated “cloud bursting”

Thanks to ...

Amazon Kurt Messersmith Iain Galvin

UCD Research IT Team Research Community Lukasz Szmit

Thank You

[email protected] http://www.ucd.ie/itservices/researchit/