Exploring The Cloud

Exploring the Cloud

Chris Sosa, Dr. Andrew Grimshaw

sosa, grimshaw @ cs.virginia.edu

University of Virginia

University of Virginia 2

Introduction - 1

Ever-increasing demand for computing resources

During non-peak times, computing resources sit idle Still paying! Power, cooling, etc

Total Cost of Ownership (TCO) is much more than the cost of Hardware Maintenance Administration Cooling Etc.


Introduction - 2

Observation – load on main ITC clusters exhibit bimodal distribution

Can we only pay for what we use?


Enter Cloud Computing (field trip!)

What is it?Infrastructure-related capabilities provided as a

serviceAlso known as utility-computing and is

associated with very basic API’sLots of industry support

Amazon Infrastructure Services: EC2, S3, … Google App EngineMicrosoft AzureIBM led initiatives


Cloud Computing Paradigms

Top-down: Client only provides program and deployment informationMicrosoft AzureGoogle App Engine

Bottom-up: Raw Infrastructure provided (virtualized hardware) AmazonNirvanixFlexiscaleGoGrid


Advantages and Disadvantages of Using the CloudAdvantages

Pay for what you use – model is based on how long you use resources. You can allocate and deallocate them on-the-fly

Hardware cost, set-up time, maintenance, cooling all go down to zero

Can start developing immediatelyDisadvantages

No control over physical resources. Do you trust Amazon? SLA’s may not be good enough. Is 99.95% availability good

enough? Some limitations in what you can run. Must stay within the API /

framework given


Why Cloud Computing

Only have to pay for what we useDisadvantages do not affect most users in a

batch system


Amazon Leading the Push

Amazon has been most successful player so farOver 29 billion objects stored on S3Using over 60% of their resources for Cloud

servicesEC2 just went out of Beta in October (new)

… rest of these slides will assume we use Amazon


Outline

IntroductionOverview of Amazon Cloud ServicesProposal of Hybrid SchedulerQuestions to be AnsweredConclusion


Amazon S3

Simple Storage for the InternetApplications can interact with various mechanisms

REST SOAP Bit Torrent

250 Mb/second network linkObjects stored in buckets

Buckets have own namespace Up to 100 buckets per account Unlimited objects per bucket 5 GB limit on size of objects Objects are write-once

SLA guarantees 99.9% availability


S3 Pricing

Storage$0.15 per GB-Month of storage used

Data Transfer$0.10 per GB - all data transfer in$0.18 per GB - first 10 TB / month data transfer out$0.16 per GB - next 40 TB / month data transfer out$0.13 per GB - data transfer out / month over 50 TBFREE to EC2

Requests$0.01 per 1,000 PUT or LIST requests$0.01 per 10,000 GET and all other requests*

* No charge for delete requests


Amazon EC2

Provides Virtual Compute Resources Purchase CPU’s on hourly

basis Can use provided virtual

machine images, or make own Virtual Machines run atop Xen

Can do meta data operations with REST, SOAP, command-line tools

Instances assigned IP address for SSH, remote desktop, etc

SLA guarantees 99.95% availability


EC2 Pricing

Instances $0.10 / hr - Small Instance - 1.7 GB of memory, 1 EC2

Compute Unit (1 virtual core - 1.7 GHz processor), 160 GB of instance storage, 32-bit platform (can buy in sets of 1, 4, 8)

$0.20 / hr - High-CPU Medium Instance 1.7 GB of memory, 5 EC2 Compute Units (2 virtual cores with 2.5 EC2 Compute Units each), 350 GB of instance storage, 32-bit platform (can buy in sets of 1 or 4)

Data Transfer $0.10 per GB – data in $0.18 per GB - first 10 TB out FREE to S3


Overview



Main Idea

Reduce the number of resources we have active and improve peak performance

Modify local schedulerWhen CPU usage is above threshold, allocate

new machines from EC2 and schedule jobsAs usage decreases, deallocate resources and

return to normal usage


Design

Amazon EC2

...Virtualized

PC...

Virtual Machine

... ...

Genesis II

... ...

Am

azon

S

3 AMI

. . .

Local Cluster

Batch Scheduler

Sch

edul

e jo

bsC

heck

usa

ge

Allocate resource on-the-flyOff-load jobs


Research Setup

Instead of spending funds on running experiments using EC2 and S3, we will be using Eucalyptus to emulate EC2Eucalyptus is an open-source implementation of the

EC2 interfaceRequires Xen be installed on host machines (need

dedicated machines)Create a centralized repository for data for our

tests (S3)NFS shareOther possibilities?


Task Bar

1. Decide on the software that will be installed on the virtual machines1. PBS licensing is complicated and expensive2. Several alternatives such as Genesis II, Hadoop, etc.

2. Create AMI image and register with Eucalyptus3. Incorporate virtual machines from Eucalyptus into existing

scheduler and create mechanism to do this on-the-fly4. Modify scheduler to take into account a threshold5. Build stubs to measure how much bandwidth, time, etc. is being

used by the scheduler so that we can determine the price we would be charged by Amazon's EC2 and S3

6. Incorporate these costs, build economic model using actual workloads at UVa, differing thresholds, and various ways of passing jobs to the Cloud


Overview



Questions to be Answered

What is the Cost Model associated with working with Cloud computing?

What costs would be associated with common jobs being run at UVa?

What software will we have installed on the Virtual Machines in the Cloud?

How can we create a threshold such that we can decide on-the-fly when to start offloading resources to Cloud resources?


Overview



Conclusions

Important to be concerned about reducing costs as well as getting bigger bang for your buck

Offloading job processing to Cloud computing infrastructures can save costs while improving peak throughput


Questions?

Exploring The Cloud

Education

amazon university of

s3 university of virginia

design university of

gb data

gb data

amazon ec2

cloud services ec2

gb of memory