Leveraging elastic web scale computing with AWS

Leveraging Elastic Web-Scale Computing
with AWS
Shiva Narayanaswamy

EC2 Basics
Instance Lifecycle
EC2 Instance Types
Using Amazon Machine Images
Bootstrapping EC2 Instances
Monitoring EC2 with CloudWatch
Autoscaling

EC2 Basics
Virtual Servers in the Cloud
• One instance to thousands of instances
• In any public AWS region
• Create, start, stop, configure, monitor as desired
• Install any software: web, business, client/server,
batch processing
• Pay only for capacity you use
• Variety of cost models Amazon EC2

EC2 Basics: cost models
On-Demand Reserved Spot Dedicated
Pay upfront in exchange for hourly
prices that are 50-75% lower than
On-Demand
Pay for compute capacity by
the hour. No long-term
commitments
Bid for unused Amazon EC2
capacity
Launch instances in VPC on
dedicated customer hardware
Customers can combine multiple purchase types to optimize pricing based on current and forecast capacity needs.
Spiky workloads Committed utilization Time-insensitive workloads Highly sensitive workloads

Provisioning and Lifecycle
• Create -> Start -> Stop -> Terminate
• Manually in console
• Automate via API (or other tools)
• Automatically based on demand
(demand curve)

Instance Types
GPU
Enabled
General
Purpose
Storage and IO
Optimized
Compute
Optimized
Memory
Optimized
M3 C3 I2
CG1M1 C1 CR1CC2 HI1 HS1
G2
M3 C3 I2 HS1
M2
R3G2
Added
Instance
Types

Amazon Machine Images
Your machine images
AMIs you have created from EC2 instances
Can be kept private or shared with other
accounts
Amazon maintained
Set of Linux and Windows images
Kept up to date by Amazon in each
region
Community maintained
Images published by other AWS users
Managed and maintained by Marketplace
partners

Bootstrapping: metadata and userdata
• Every EC2 Instance has access to local instance
metadata and userdata service
Instance
request
User
data
Instance
Meta-data
service

• Metadata: immutable information about the instance
• Accessible from within the instance via HTTP at
http://169.254.169.254/latest/meta-data/
• Script(s) on instance may retrieve useful information about the
instance, such as:
• Host name
• AMI ID
• Instance ID
• Public/Private DNS
• Availability Zone

• User Data: pass up to 16KB of text to an instance on launch
• Accessible from within the instance via HTTP at
http://169.254.169.254/latest/user-data/
• Text can be parsed by script on instance and used to configure the
machine

Custom script on AMI
(script_runner.py) fetches userdata,
parses it, and configures EC2 Instance
on boot

• CloudInit executes UserData on first boot if UserData begins with:
• #! (Linux)
• <script> (Windows; technically, EC2Config, not CloudInit, does this)
• CloudInit is installed on Amazon Linux, Ubuntu, and RHEL AMIs
• EC2Config is installed on Windows Server AMIs
• Both may be installed on other distributions via a package repo or
source
Bootstrapping: UserData and CloudInit

• UserData to install Apache and MySQL on boot, and attach an EIP:
#!/bin/bash
# Install Apache, PHP, and MySQL
yum install –y httpd mysql-server
# Attach an Elastic IP to this instance
ec2-associate-address
23.34.45.56
-i $(curl http://169.254.169.254/latest/meta-data/instance-id)
Bootstrapping: UserData and CloudInit

Bootstrapping
Bake an AMI
Start an instance
Configure the instance
Create an AMI from your
instance
Start new ones from the AMI
Configure dynamically
Launch an instance
Use metadata service and
cloud-init to perform actions on
instance when it launches
Use config management tools
like Puppet/Chef/Opsworks
vs

Bootstrapping
Bake an AMI Configure dynamically
Build your base images and
setup custom initialisation
scripts
Maintain your ‘golden’ base
Use bootstrapping to pass
custom information in and
perform post launch tasks.
+
Sweet spot

Bootstrapping: AMIs
Linux
JEE
Your Code
Log4J
Spring
Hibernate
Struts
Tomcat
Apache
Java App Stack
Example full stack required to run your
application.
Let’s use the 3 bootstrapping
techniques

Bootstrapping: AMI bake
Fully-functional AMI is pre-build and
ready to launch from the AMI inventory
Inventory of AMIs
Linux
JEE
Your Code
Log4J
Spring
Hibernate
Struts
Tomcat
Apache
Amazon EC2
Linux
JEE
Your Code
Log4J
Spring
Hibernate
Struts
Tomcat
Apache
Linux
JEE
Your Code
Log4J
Spring
Hibernate
Struts
Tomcat
Apache
Linux
JEE
Your Code
Log4J
Spring
Hibernate
Struts
Tomcat
Apache
Linux
JEE
Your Code
Log4J
Spring
Hibernate
Struts
Tomcat
Apache
Java AMI

Bootstrapping: Configure dynamically
Base OS AMI
An AMI with minimal components (OS,
J2EE, and Chef/Puppet) is launched.
All configuration occurs via
Chef/Puppet after instance launch
Inventory of AMIs
Amazon EC2
OS AMI
Fetch on boot
Linux
JEE
Your Code
S3
Hibernate
Tomcat
Log4J
Spring
Struts
Apache
Linux
JEE
Linux
JEE
Chef/Puppet
Chef/Puppet
scripts

Bootstrapping: Sweet spot
Partially-configured AMI
A “Golden Image” is launched, with
scripts fetching/installing app code
and other supporting components on
boot
Inventory of AMIs
Amazon EC2
Java AMI
Your Code
S3
Log4J
Spring
Struts
Linux
JEE
Hibernate
Tomcat
Apache
Fetch on boot
Fetch on boot
Linu
x
JEE
Hibe
rnat
e
Tom
cat
Apac
he
Linu
x
JEE
Hibe
rnat
e
Tom
cat
Apac
he
Linu
x
JEE
Hibe
rnat
e
Tom
cat
Apac
he
Linu
x
JEE
Hibe
rnat
e
Tom
cat
Apac
he

Why do this?
Automation
Less fingers, less mistakes
Availability
Drive higher
availability with self-
healing
Security
Instances locked
down by default
Flexible
Shell,
Powershell,
CloudFormation,
Chef, Puppet,
OpsWorks
Scale
Manage large scale
deployments and drive
autoscaling
Efficiency
Audit and manage
your estate with
less time & effort

Do Don’t
Some dos and don’ts
Use IAM roles
Go keyless if you can
Strike a balance between AMI and
dynamic bootstrapping
Put your API access keys into code
(and then publish to GIT) or bake
into AMIs (and share)


Monitoring EC2 with CloudWatch

Types of Scaling
• Vertical Scaling
• Changing instance size
• Increasing EBS Capacity
• Horizontal Scaling
• Adding / removing instances
• ELB
• Autoscaling
r3.8xlarge
c3.2xlarge
m3.medium
m3.medium m3.medium m3.medium
m3.medium m3.medium m3.medium

Vertical Scaling
• Different EC2 instance type
• High memory instances
• High CPU instances
• High I/O instances
• High storage instances
• Easy to change instance sizes
• Will hit an endpoint eventually
• Requires instance to be stopped
r3.8xlarge
c3.2xlarge
m3.medium

Traditional IT Usage Patterns
On and Off Fast Growth
Variable peaks Predictable peaks

Traditional IT Usage Patterns
Poor
Service
WASTE

Cloud IT Usage Patterns (Auto Scaling)

Auto Scaling
• Automatic resizing of compute clusters based on demand
• Define minimum and maximum number of instances
• Define when scaling out and in occurs
• Use metrics collected in Amazon CloudWatch to drive scaling
• Run Auto Scaling for On-Demand and Spot instance types
• Its Free! Amazon
CloudWatch
Usage
Metrics
Scaling
Instructions
Auto Scaling Group
Queue
Metrics
AutoScaling

Describes what Auto Scaling
will create when adding
Instances - Similar to ec2-run-
instances API command
AMI
Instance Type
Security Group
Instance Key Pair
Only one active launch
configuration at a time
Auto Scaling will terminate
instances with old launch
configuration first
rolling update
Auto Scaling managed
grouping of EC2 instances
Automatic health check to
maintain pool size
Automatically scale the number of
instances by policy – Min, Max,
Desired
Automatic Integration with ELB
Automatic distribution &
balancing across AZs
Parameters for performing an
Auto Scaling action
Scale Up/Down and by how much
ChangeInCapacity (+/- #)
ExactCapacity (#)
ChangeInPercent (+/- %)
Cool Down (seconds)
Policy can be triggered by
CloudWatch events
Launch Configuration Auto-Scaling Group Auto-Scaling Policy

Scaling plan
• Scale based on demand
• Manual scaling
• Scale based on schedule
• Maintain current instance levels at all time
AutoScaling

Availability Zone A Availability Zone B
Autoscaling: Auto Scaling Group

Latency
CloudWatchAuto Scaling
ELB
Auto scaling Group
Autoscaling: ELB + CloudWatch

• Tools Used:
• CloudFormation script –
• Create a multi-AZ, load balanced and Auto Scaled sample web site running on an Apache
Web Server (m1.small). The application is configured to span all Availability Zones in the
region and is Auto-Scaled based on the CPU utilization of the web servers.
• Bees with Machine Guns – Performance testing tool
• A cloudformation script that spins up a distibuted performance testing tool based on
apache eb tool. This tool will hit the ELB with 1000’s of concurrent requests for a total of
100’s of thousands of request, thus loading the web server behind the ELB.
• Expected result
• The Apache web server will scale to serve traffic without any customer impact.
Autoscaling: DEMO

• CloudFormation script (Auto scaling apache web server)
• Auto-scaling group configuration:
• Min: 1
• Max: 3
• Cooldown: 300
• Scaling Policies:
• Scaling Up:
• CPU Utilization > 20% for 1 consecutive period of 60 seconds
• Action: Add 1 instance
• Then wait: 60 seconds before next operation
• Scaling Down:
• CPU Utilization < 10% for 2 consecutive periods of 60 seconds
• Action: Remove 1 instance
• Then wait: 60 seconds before next operation
• Bees with Machine guns(NASTY)
Demo Information

Autoscaling isn’t one size fits all
• Choose the right metrics
• CPU Usage
• Queue Depth
• Number of concurrent users
• Scale too aggressively
• Overprovisioning: increases costs
• Bounciness: Add more than you need and have to partially scale back shortly after
scaling up, increasing costs.
• Scale too timidly
• Poor performance
• Outages due to lack of capacity
• Scale out early / Scale in slowly

Stop doing these:
Provisioning and fixing servers
Treating compute as physical things
Thinking of compute as a finite commitment

and start doing these
Security
Build systems secure by
default
Elasticity
Stateless autoscaling
applications
Replace not fix
Build from scratch, don’t
fix something
Unconstrained
Say goodbye to
traditional capacity
planning
Be cost aware
Tag resources, play with
instance types
Automation
Create instances when
you need them, drop
them when not

What’s more?
• Attach / Detach Instances from Auto Scaling Groups
• Place instances into Standby State to Troubleshoot
• Hold instances in Pending state for installing software / retrieve logs
• Create an Auto Scaling Group / Launch Configuration based on a
running instance
• Auto scaling Lifecycle hooks

Leveraging elastic web scale computing with AWS

Recommended

More Related Content

Viewers also liked (11)

Similar to Leveraging elastic web scale computing with AWS (17)

More from Shiva Narayanaswamy (12)

Recently uploaded (20)

Leveraging elastic web scale computing with AWS