Optimizing Costs and Efficiency of AWS Services

Chicago

©2015,  Amazon  Web  Services,  Inc.  or  its  affiliates.  All  rights  reserved

Enterprise Pre-Day

Do you • Wear green eye-shades? • Work from the console?

You’re using AWS. You like it.

But maybe you are spending more than you planned

Or you’d just like to spend less

What should you do?

How about some

frameworks

Some tools

Some best practices

And some examples from other customers

like you

How are you controlling the provisioning of

resources?

Control who can provision resources

Require tagging

“No tags? No instance” – Large financial services customer in Boston

Best practices: • Control who can provision

resources via IAM • Require tagging

– Inspect – Stop instances without tags

Savings potential

100% if you don’t start instances you don’t need!

Are you identifying and terminating idle

resources?

How do I identify idle resources?

AWS Trusted Advisor runs 100+ configuration

checks and recommends savings

Trusted Advisor Demo

We are using AWS Trusted Advisor to improve cost efficiency and to audit

the configurations of service platforms

Trusted Advisor

How do I handle demand spikes?

Design for elasticity rather than deploy for

peak

Use Auto Scaling or queue-based approaches

to add resources when needed, and turn them

off when not

Amazon SQS Queue

Auto Scaling

Problem: Workloads arrive asynchronously and require isolation

Solution: Amazon

Simple Queue Service and launch/

termination of instances

Amazon SQS

Problem: How to scale up to accommodate sudden traffic

spikes?

Solution: Auto Scaling Auto Scaling

Auto Scaling real world example •  Baseline traffic: 300+ thousand pages / hour

130 TB/month outbound traffic •  Peak: 2.0+ million pages / hour (~7x baseline)

•  Architectural changes: e.g., move from local SSDs to provisioned IOPS needed due to this scale

•  Auto Scaling provides 23% savings over three years

Best practices •  Identify idle resources using Trusted

Advisor •  Turn them off

•  Design to scale up and scale down using Auto Scaling or a queue-based approach

•  Match your usage and capacity — don’t pay for idle resources!

Are you using the most cost-effective

instances or resources?

Match resources to

workload

How can I tell if there’s a match or mismatch?

Use Amazon CloudWatch to

collect and track metrics

CloudWatch

Amazon CloudWatch monitoring Basic •  7 metrics for Amazon EC2

including: –  CPU utilization –  Data transfer –  Disk usage and more

•  5-minute frequency •  Metrics for Amazon EBS,

Amazon DynamoDB, Amazon RDS, etc.

Detailed •  1-minute frequency •  Aggregation by instance type and

AMI

Why does this matter?

Different instances cost different amounts

Instance families: Example Instance vCPU Mem

(GiB) Monthly price (3-yr AURI )

Ideal use case

c4.2xlarge 8 15 $125.44 Best price-compute performance

m4.2xlarge 8 32 $137.25 Balanced

r3.2xlarge 8 61 $179.25 Lowest cost per GiB RAM

r3.xlarge 4 30.5 $89.61 Lowest cost per GiB RAM

Do you need continuous compute

or does bursting work for you?

t2.medium •  4 GiB RAM •  Baseline performance: 40% •  Bursts beyond this based on CPU

credits •  Less than $17 per month •  Details: http://amzn.to/1sl2bKa •  Web servers, dev, small databases

t2

m3.medium

t2.medium (bursting)

Cloudwatch: CPU Utilization

t2 vs m3.medium or c4.large Instance vCPU Mem

(GiB) Monthly price (3-yr AURI )

Ideal use case

m3.medium 1 3.75 $19.08 Always available, balanced

c4.large 2 3.75 $31.36 Always available, compute

t2.medium 2 4 $16.86 Bursty workloads

Consider switching to a t2?

Guidance •  Inspect your workloads (use CloudWatch!) •  Can your workload run on a t2? •  If not, is it memory intensive? àr3 •  Is it compute intensive? à c4

•  Others to consider: –  i2 for storage-optimized –  g2 for GPU

Savings potential If t2 works for you •  11%: Switch from m3.medium to t2.medium •  46%: Switch from c3.large to t2.medium Instance optimization •  9%: Switch to c4 for compute-intensive apps

(m4.2xl -> c4.2xl) •  35%: Switch to r3 for memory-intensive apps

(m4.2xl -> r3.xlarge)

Best practices: 1. Monitor instances with CloudWatch 2. Switch to t2 as possible 3. Use compute- and memory-focused instances as needed

Do you need long-term storage?

Amazon Glacier Extremely low-cost cloud

storage service

Amazon Glacier

Are you purchasing in the most cost-

effective way?

Commitment and spot save $

Use Reserved Instances if possible:

Up to 60% savings + capacity reservation

AWS Trusted Advisor identifies savings

attainable via Reserved Instancess

Reserved Instance Marketplace

Use Spot Instances for non-stateful workloads

Pricing starts

at 90% off

Novartis, Cycle Computing & AWS Spot Instances

Goal: Screen 10 million compounds

against a common cancer target in < 1 week

10,600 AWS instances

39 drug design years in 11 hours

3 promising compounds identified

Cost of $4,232

Savings potential • ~40%: 1-Year Reserved

Instances • ~60%: 3-Year Reserved

Instances • Up to 90%: Spot Instances

Could you save by letting AWS do more

work for you?

Focus on what you do best

Let AWS solve other problems for you

AWS’s higher-level services automate your work and save

you money CloudFront

DynamoDB

Amazon RDS

ElastiCache

Amazon Redshift

Amazon EMR

Amazon Kinesis

Amazon WorkSpaces

Are you modeling and monitoring your

spend?

https://calculator.s3.amazonaws.com/index.html

Detailed billing reports and

cost insights

Detailed Billing Reports http://amzn.to/1swNwLV

Cost Explorer http://amzn.to/1zHE2Fj

Cost Explorer Demo

Forecasting Demo

Budgeting Demo

Demo

CloudWatch

•  Control and verify •  Design for elasticity •  Match resources and

workload •  Purchase for savings •  Use application

services •  Model and track

Frameworks to remember

Tools to use •  IAM, AWS console •  Auto Scaling,

Amazon SQS •  Amazon CloudWatch •  AWS Trusted Advisor •  Detailed billing reports

and Cost Explorer •  Partners

•  Turn off untagged resources •  Stop idle instances •  Use instances matched to task

and less expensive ones •  Baseload with RIs •  Use Spot Instances for non-

stateful workloads •  Model and track spending

Best practices

Chicago

©2015,  Amazon  Web  Services,  Inc.  or  its  affiliates.  All  rights  reserved

Enterprise Pre-Day