Optimizing Total Cost of Ownership for the AWS Cloud

© 2014 Amazon.com, Inc. and its affiliates. All rights reserved. May not be copied, modified, or distributed in whole or in part without the express consent of Amazon.com, Inc.

Optimizing Total Cost of Ownership

for AWS

Rohit Rahi, Amazon Web Services

Valentino Volonghi, AdRoll

Agenda

• Total Cost of Ownership

• AdRoll

• Cost Optimization on AWS

Lower costs with AWS

1

“Average of 400 servers

replaced per customer”

Replace up-front

capital expense with

low variable cost

2

42 Price

Reductions

Economies of scale

allow AWS to continually

lower costs

3

Pricing model choice

to support variable &

stable workloads

4

Save more money as

you grow bigger

Tiered Pricing

Volume Discounts

Custom Pricing

On-Demand

Reserved

Spot

Dedicated

Source: IDC Whitepaper, sponsored by

Amazon, “The Business Value of Amazon

Web Services Accelerates Over Time.”

December 2013

Lower costs than on-premises

On-Premises

Traditional

Data Center

On-Premises

Virtualized

Data Center

CAPEX

OPEX

OPEX

AWS

CAPEX

OPEX*

Cost savings from running

internal IT more efficiently AWS Scale• Multiple new data centers built each year

• Volume purchasing, highly automated, supply

chain optimization

Utilization fundamentally higher in

AWS cloud• Aggregating non-correlated workloads, scale,

spot market

Amazon specific hardware designs• OEM acquisition of custom servers & net gear

• Direct purchasing of disk, memory, & CPU

• AWS controlled hypervisor & net protocol

layers

Diagram is not to scale

*For AWS, OPEX costs includes Reserved Instances one-time low, upfront payment, if Reserved Instances are used.

Cost savings from moving

to a public cloud provider

AWS Pricing Philosophy

More AWS Usage

More Infrastructure

Economies of Scale

Lower Infrastructure

Costs

Reduced Prices

More Customers Ecosystem

Global Footprint

New Features

New Services

Infrastructure

Innovation

We pass the savings along to our

customers in the form of low

prices and continuous reductions

42

Analysts have shown AWS reduces costs

In early 2012, AWS commissioned IDC to interview 11 organizations that deployed applications on AWS. Since this study was conducted in early

2012, AWS has introduced price reductions nearly 20 times across Amazon EC2 and Amazon S3. IDC estimated what the impact of AWS's fee

restructuring would be on the organizations that participated in the 2012 study and determined that the overall fees would drop by 21% lowering the

five year TCO from $909,000 to $846,000. Source: IDC Business Value of AWS Accelerates over time

IT

PRODUCTIVITY

INCREASE:

52%

5 YEAR TCO

SAVINGS:

72%

AWS TCO benefits increase over time…

$3.50 in benefits

$1 Investment in AWS $1 Investment in AWS

$8.40 in benefits

At 36 Months of using AWS… At 60 Months of using AWS…

~3X ~8X

Source: IDC Business Value of AWS Accelerates over time

According to IDC, this relationship between length of time using AWS and return is due to customers leveraging the more optimized

environment to generate more applications along a learning curve.

Comparing TCO is not easy(But We’re Going to Try)

≠

Typical cost drivers for on-premises deployments,

including overhead costs

Network

Costs

Storage

Costs

Server

Costs

Hardware – Server, Rack

Chassis PDUs, ToR

Switches

(+Maintenance)

Software - OS,

Virtualization Licenses

(+Maintenance)

Overhead Cost

Space Power Cooling

Hardware – Storage

Disks, SAN/FC Switches

Overhead Cost

Storage Admin costs

Network Hardware – LAN

Switches, Load Balancer

Bandwidth costs

Network Admin costs

Overhead Cost

IT Labor

Costs

Server Admin

Virtualization Admin

1

2

3

4

Space Power Cooling

Space Power Cooling

illustrative

Diagram doesn’t include every cost item. E.g. software costs can include database, management, middle tier software costs.

Facilities cost can include costs associated with upgrades, maintenance, building security, taxes etc. IT labor costs can include

security admin and application admin costs.

AWS services pricing includes overhead costs

Hardware

Vendor

Offering

✔ ✔ ✔ ✔ ✔ ✔ ✔ ✔

✔

ServerNetwork Hardware

Software OS +VMs

DC/Co-loFloor Space

PoweringCooling

Software Defined

Networking

Data CenterPersonnel

Storage Redundancy

Resource Mgmt. /SWAutomation

× × × ××× ×

TCO Example : 100 VMs On-Premises vs. AWS

# of VMs Avg.

vCPU

Avg.

vRAM

Optimize by? Usage

25 1 2 RAM 5%

35 4 14 RAM 80%

30 8 32 RAM 60%

5 8 68 RAM 40%

5 16 128 Disk I/O 8%

# of

InstancesInstance vCPU RAM Instance Type

25 m1.small 1 1.7 On Demand

35 m3.xlarge 4 15 3 Yr. Heavy RI

30 m3.2xlarge 8 30 3 Yr. Med. RI

5 m2.4xlarge 8 68.4 3 Yr. Light RI

5 i2.4xlarge 16 122 3 Yr. Light RI

Avg.

vCPU

Avg.

vRAM

Optimize by?

1 2 CPU

4 14 CPU

8 32 CPU

8 68 CPU

16 128 Disk I/O

Instance vCPU RAM

m1.small 1 1.7

c3.xlarge 4 7

c3.2xlarge 8 15

c3.2xlarge 8 15

i2.4xlarge 16 122

1, 1.7, $0.060

1, 3.75, $0.113

2, 3.75, $0.145

2, 7.5, $0.225

2, 17.1, $0.410

4, 7, $0.300

4, 15, $0.450

4, 34.2, $0.820

8, 15, $0.600

8, 30, $0.900

8, 68.4, $1.640

4, 30.5, $0.853

8, 61, 1.705

16, 30, $1.200 32, 60, $2.400

32, 244, $3.500

16, 122, $3.410

16, 117, $4.600

32, 244, $6.820

0

50

100

150

200

250

300

0 5 10 15 20 25 30

TCO Example : Choosing EC2 instances

On Demand Prices shown (N.Virginia region), only latest generation instances (M3,C3) shown

where applicable, GPU and Micro instances not shown above

Memory-Optimized Instances

Compute-Optimized Instances

General Purpose Instances

Storage-Optimized Instances

vCPU

RA

M

TCO Example : 100 VMs On-Premises vs. AWS

• From over 40 data centers down to 6

• Planning to migrate 3000 apps by Jan

2015

• Saving $100M over 3 Years

VS

1. Evaluate infrastructure

costs & architecture

2. Make business case 3. Enable decision to move to the

cloud

Customer Spotlight: Dow Jones Intl.

In Your TCO Analysis

Power/Cooling (compute, storage, shared network)

Data Center Administration (procurement, design, build, operate, network,

security personnel)

Rent/Real Estate (co-lo charges, building deprecation, taxes)

Software Licensing/Maintenance (OS, Virtualization, DCIM, Automation..)

RAW vs. USABLE storage capacity (Usable = ~50% Raw)

Storage Redundancy (RAID penalty, OS penalty)

Storage Backup costs (Tape, backup software)

Bandwidth, Network Gear & Redundancy (Routers, VPN, WAN..)

DON’T

FORGET

THINK

BENEFITS

Additional investment into new initiatives

Reduced Procurement Time, Resource sitting on shelf

Cost of lost customers

Lower down time, increased productivity

50B req/day without breaking the bank

Pixel “fires”

Pixel “fires”

Serve ad?

Pixel “fires”

Serve ad?

Ad served

Data must be available all over the world!

7/2011 - ~50GB/day

4/2013 - ~5TB/day

10/2013 - ~20TB/day

03/2014 - ~40TB/day

What do we do then?

• Store 1.5+ PB of compressed data in S3

• Be available worldwide, and replicate all our data in every major

area

• ~50B requests/day with 100ms latency caps

• As close as possible to 100% uptime

• 35 engineers

AdRoll S3 Usage by the numbers

• 1.5+ PB of compressed data stored

• 5-10TB of compressed new data each day

• ~5B monthly requests in February

• 500+ TB of transfer

• No engineers spend time on storage

AdRoll S3 buckets would be one of the most

visited properties on the web, we don’t even

think about it.

BackBlaze Storage Pod 4.0

• 180TB cost $9,305 to build, only 1 PSU per box.

• X 3 for local redundancy

• X 3 for geographic redundancy

• X 2 because we need to have room for growth

• +50% because we need to have quickly available spare parts in

each area

• TOTAL: $251,235 <- this is the real cost of 180TB

And it’s not even all of it…

• Now let’s develop the software that manages about 5B monthly

requests and transparently fails over…

• Let’s add bandwidth cost, electricity and data center costs…

• Our AWS total cost is equivalent to about 6 engineers.

We’re not a storage company!

We happen to have a similar amount of

data that a storage company has. But

we don’t have engineering dedicated

to storage.

Our RTB infrastructure

• ~500 c3.4xlarge in 4 regions

• Close to 100% uptime since launch

• <100ms max latency, 0.15% timeouts YTD at 50B requests/day

• 500K+ requests/second globally on DynamoDB

• Fully replace the global infrastructure in less than 1 hour

2 engineers

We have more people installing,

configuring and managing our office

network and windows laptops

Alternative?

• Setup 4 data centers locations

• 2 on-call staff in each location

• Provision ~1000 machines for growth and deployment flexibility.

• Implement Auto Scaling and a common API

• Provision 10-20% cold capacity in each location

• What if a big customer signs up for AdRoll?

• What if we improve our algorithms and we can do with half the machines? What if we want to upgrade our hardware?

We’ve been trading money… For Time

Time is a far more scarce resource and

you can’t save it.

What does it mean in numbers?

• Minimum team size is about 20:

– 8 on-call

– 1 product manager for Auto Scaling and API

– 5 engineers to develop and maintain Auto Scaling

– 5 engineers to maintain Cassandra installation instead of DDB

– 1 engineering manager

• 20 people at an average $130K/year…

$2.6M/year

We spent less than that to run our entire

infrastructure in EC2 last year.

All of these very skilled engineers would

not be adding value to our product.

If we didn’t use EC2 a competitor

would…

And then we’d be in trouble

Cost Optimization on AWS

1. Choose the right instance types

Start

Choose an instance

that best meets your

basic requirements

Start with memory & then

choose closest virtual

cores

Look for peak IOPS

storage requirements

Tune

Change instance size up

or down based upon

monitoring

Use CloudWatch &

Trusted Advisor to assess

Roll-Out

Run multiple instances

in multiple Availability

Zones

2. Use Auto Scaling

Describes what Auto Scaling will

create when adding Instances

Only one active launch configuration

at a time

Launch

Configuration

as-create-launch-config

--image-id ami-54cf5c3d

--instance-type m1.small

--key mykey

--group webservers

--launch-config 101-launch-config

Auto Scaling managed grouping of EC2

instances

Automatically scale the number of

instances by policy – Min, Max, Desired

Auto Scaling

Group

as-create-auto-scaling-group 101-as-group

--availability-zones us-east-1a us-east-1b

--launch-configuration 101-launch-config

--load-balancers myELB

--max-size 5

--min-size 1

Parameters for performing an Auto

Scaling action

Scale Up/Down and by how much

Auto Scaling

Policy

as-put-scaling-policy 101ScaleUpPolicy

--auto-scaling-group 101-as-group

--adjustment=1

--type ChangeInCapacity

--cooldown 300

Utilization and Auto Scaling: Granularitymore small instances vs. less large instances

29 m1.large @ $0.240/hr.

= $6.96

59 m1.small @ $0.06/hr.

= $3.54

3. Turn off un-used instances

• Dev./test instances

• Simple instance start/stop

• Tear down/build up altogether

• Instances are disposable

0

2

4

6

8

10

12

14

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

Light Utilization RI

Medium Utilization RI

Heavy utilization RI

/Spot Instances

4. Use Reserved Instances

Reserved Instances enable you to maximize savings by paying a low, one-time fee for a

capacity reservation of 1-3 years in exchange for a significant discount on the hourly rate.

Amazon EC2up to 65% savings

Amazon RDSup to 76% savings

Amazon DynamoDBup to 76% savings

Amazon Redshiftup to 73% savings

Amazon ElastiCacheup to 70% savings

AWS services offering reservations

Customer Spotlight: Pinterest

AWS

Set-up

• 8 billion objects and 400 terabytes of data (May 2012), 10x growth from August 2011; EC2 instances

have grown by 3x in the same time period

• 150 EC2 instances ( web tier), 90 instances (in-memory caching), 70 master databases

• Reserved instances used for standard traffic; On-demand and spot instances used to handle daily elastic

load. Most services targeted to run at about 50% on-demand and 50% spot

Source: Pinterest AWS Case study, Pinterest Architecture update, 410TB of data, 10X Growth, 12 employees

Source: Return on Agility, Werner Vogels

“Imagine we were running our data center, and we had to go through a process of capacity

planning and ordering and racking hardware. It wouldn't have been possible to scale fast

enough,"– Ryan Park, Pinterest Operations Engineer

Costs have gone from $54 per hour to $20

per hour

Only 2 weeks of engineering effort was

required to achieve this cost savings

5. Use Spot Instances

• Pricing

• Up to 92% discount

• Elastic

• Capacity not otherwise

available

• Tradeoff

• Potential for interruption

Picking the right Bid Price - Tolerance for interruptions, % likelihood of

termination

1.21 PFLOPS

264 years of compute in < 18 hours

16,788 instances in 8 regions

Customer Spotlight: Cycle Computing

On-Premises Spot

$68 Million

$33K

6. Leverage Storage Classes

AWS Cloud

Amazon

Glacier

Gateway Appliance/

AWS Storage Gateway

Amazon

S3

Block File

On-premises Data Center

Archive Backup Disaster

Recovery

Amazon

EBS

Amazon S3 Reduced Redundancy

• 99.99% durability vs. 99.999999999%

• Up to 20% savings• Great for everything that is easy to

reproduce

Amazon Glacier

• Same durability as S3• 3 to 5 hours restore time• Up to 89% savings• Great for archiving, long-term

backups and old data

7. Offload your architecture

The more you can offload, the less infrastructure you need to maintain, scale, and pay for

• Offload popular traffic to Amazon CloudFront and S3

• Introduce Caching

Re

spo

nse

Tim

e

Se

rve

r L

oa

d

Re

spo

nse

Tim

e

Se

rve

r

Load

Re

spo

nse

Tim

e

Serv

er

Load

No CDN CDN for

Static

Content

CDN for

Static &

Dynamic

Content

8. Use Application Services

Elastic

Load

Balancing

Amazon

Relational

Database

Service (RDS)

Amazon Simple

Queue Service

(SQS)

Amazon Simple

Email Service

(SES)

Amazon

Elastic

MapReduce

Amazon

ElastiCache

Amazon Simple

Notification

Service (SNS)

Web Servers

Availability Zone

$0.025 per Elastic Load Balancer-

hour (or partial hour)

$0.008 per GB of data processed

by an Elastic Load Balancer

100 GB Data processed, 1 ELB

$18 (.025*24*30) + $.008*100

$18.80

Web Servers

Availability Zone

EC2 instance

+ software LB

Elastic Load

BalancerDNS

DNS

VS

Leverage Application Services

$0.060 per hour, m1.small

Separate for Software Load

Balancer

$.060*24*30 = $43.2 (m1.small)

+ Software LB Cost

On Demand Prices shown (N.Virginia region)

9. Use Consolidated Billing

• Receive a single bill for all charges incurred across all linked accounts• Share RI discounts

• Combine tiering benefits

• View & manage linked accounts

• Add additional accounts

10. Leverage AWS tools

AWS Trusted Advisor

AWS EC2 Usage Reports

Recap

Choose the right Instance type

Use Auto Scaling

Turn off un-used Instances

Use Reserved Instances

Use Spot Instances

Leverage Storage Classes

Offload your architecture

Use Application Services

Use Consolidated Billing

Leverage AWS Tools – Trusted Advisor, EC2 Usage Reports

Others…

Summary

TCO

• Make reasonable assumptions and leverage industry benchmarks

• Know the on-premises hidden costs

Cost Optimization

• Create cost-aware architectures and leverage best practices

• Re-evaluate and revisit your architecture often

• Leverage Application Services, CloudWatch

• Stay up to date – RI modifications, Trusted Advisor

© 2014 Amazon.com, Inc. and its affiliates. All rights reserved. May not be copied, modified, or distributed in whole or in part without the express consent of Amazon.com, Inc.

Optimizing Total Cost of Ownership

for AWS

Rohit Rahi, Amazon Web Services

Valentino Volonghi, AdRoll

Thank you!