High Performance Web Applications

High Performance Web Applications matthew@amazon.com Dr. Matt Wood @mza

Building blocks for wonderful applications Storage, compute, databases, managed services

Retail Merchant services Web services

Getting started with Amazon Web Services quick review H

Utility computing Available at the end of an API call.
Metered billing. H

Family of 20 services Adding new services and reﬁnement weekly.
Rapid rate of innovation. H

Undi!erentiated heavy lifting Focus on your app, customers and business
H

aws.amazon.com H Free account. Free tier.

Building High Performance Web Applications r

Deliver a great experience to your customers. Responsive, fast web
applications. r

g DECOUPLED SCALE OUT S OPTIMIZE COST FOR C AUTOMATE
v CALIBRATED FOR: CPU performance v CALIBRATED FOR: IO performance 5patterns for performance

g DECOUPLED

Monolithic Mega System v2

Application Assets Analytics

Images, javascript, stylesheets, HTML pages Deliver as fast as possible
to customers r

Websites without the servers Highly durable: 99.999999999% Zero conﬁg: ﬁre
and forget r

Low latency delivery of assets Local caching for fast downloads
Amazon CloudFront r

Create content distributions From objects in S3 Any custom origin
Dynamic, streaming and static content r

CNAME? Map to any custom URL assets.domain.com, www.domain.com r

Application architecture Small things, loosely coupled r

Load balancer App servers Database Asynchronous processing Application Assets Analytics

Shared responsibility Infrastructure by AWS OS, application and data by
customers Tools to help meet best practices r

Security groups Customer controlled ﬁrewalls All ports are closed by
default r

x

Decoupling provides an architectural framework for performance Easier to separate
concerns: assets, streaming, etc. Easier to manage and scale. r

SCALE OUT

Elasticity is a tenet of the cloud Scale up when
you need to Scale down when you don’t r

Predictable peaks r Time Demand Capacity

Variable usage r Time Demand Capacity

Variable usage r Time Demand

Horizontal scaling drives performance up ...and costs down r

Amazon Machine Images encapsulate your app Customize your virtual machines
AMI can be the unit of deployment r

1. Spin up a new instance using your AMI Conﬁgured
to bootstrap the app. r

2. Pull down latest code base From S3 or code
repository r

3. Start up the app on the instance Ready to
receive requests r

4. Register with the load balancer Job done. r

Horizontal scaling allows for fault tolerance r

Load balancer App servers Database Asynchronous processing :( Application Assets
Analytics

Deploying and scaling a datastore r Follow the same patterns:
horizontal scale, availability, automation.

One question: r Does your application require a strict, controlled
schema for query ﬂexibility?

Amazon Relational Database Service r Management systems: CRM, ERP, ﬁnance
Does your application require a strict, controlled schema for query ﬂexibility? Yes

Focus on your app r Handles tedious database admin tasks
Designed for availability

Multi-engine r MySQL, Oracle, Microsoft SQL Server Up and running
in six clicks

Point in time snapshots r Automatic. Easy recovery.

High availability r Deployed across multiple availability zones. Synchronous writes.

Read replicas r Perfect for read heavy applications. Asynchronous writes.
Replication metrics available.

Amazon DynamoDB r Web apps, social apps, mobile apps, user
generated content, unstructured data integration, lots of data. Does your application require a strict, controlled schema for query ﬂexibility? No

Focus on your app r Managed NoSQL database service. No
schema.

Unlimited scale r Unlimited storage Pay as you go

High performance r Single digit millisecond latencies

Zero admin r No instances to manage Tiny API, perfectly
formed

Asynchronous processing r Resize images. Transcode movies. Resample media.

Follow the same best practices r Decouple. Stateless. Horizontal scale.

CloudFront for dynamic content r Edge caching for dynamic content
Cache by query string parameters Multiple origin servers Persistent connections to origin servers

C AUTOMATE

Automate everything The tedious. The time consuming. The error prone.
r

For example: when increasing capacity Launch instance. Update application code.
Register with load balancer. r

Auto-scaling CloudWatch Auto-scaling service Elastic Load Balancer r

Set operational thresholds Network capacity > 80% Memory utilization >
75% CPU < 50% r

Automatically respond to operational alarms r Network capacity > 80%:
add instances Memory utilization > 75%: add instances CPU < 50%: remove instances

Set bounds Step change conﬁgured by number or percentage of
instances r

Custom metrics Respond to custom metrics Page load time Database
query time r

Automatic load balancer conﬁguration Instances added/removed automatically Health check for
running application before routing tra"c r

Stateless operations are important Decoupled, but stateless Instances are naive
to other instances r

Treat your data as a royal garden Backup to S3:
high durability Compute is a disposable resource r

v CALIBRATED FOR: CPU performance

Range of resources Choose the best instance type for performance
r

r t1.micro m1.small cc2.8xlarge

r t1.micro m1.small cc2.8xlarge High memory High CPU High CPU
+ memory

1 instance for 100 hours = 100 instances for 1
hour

Balance cost/performance Mix instance types for application servers, databases, ancillary
services, analytics etc. r

Migrate to large instance types As your application grows Use
64 bit from the outset, if possible r

v CALIBRATED FOR: IO performance

IO bound web applications are more common Application requires predictable,
consistent IO r

Provisioned throughput with DynamoDB No need to provision storage Provision
IO requirements: reads and writes per second r

Scale up, and carry on DynamoDB will manage resources to
achieve and maintain throughput r

Low latency, unlimited scale Writes acknowledged when saved to disk
in multiple data centres. Backed on SSDs. r

...and SSDs for all

hi1.4xlarge instances High capacity, high performance storage 2 x 1Tb
SSD drives r

Perfect match for databases Relational and NoSQL storage r

Perfect match for databases Relational and NoSQL storage 90k to
120k IOPS r

High performance network 10 gigabit ethernet Placement groups Perfect for
distributed stores r

S OPTIMIZE COST FOR

Economies of scale Drive down prices for customers 20 price
drops in past 6 years r

Maximize utilization Greater utilization drives down cost r

Reserved capacity Time Utilization

Reserved instances Small upfront payment Guaranteed capacity for 1 or
3 years Lower hourly rate r

Reserved instance resale Sell your reservations on the resale market
Sellers set your own price Buyers select and check out r

Reserved capacity Time Utilization

Reserved capacity On-demand Time Utilization

Spot instances Name your price computing r

Spot instances Lower costs ‘Overclock’ your app: more bang for
your buck r

Load balancer App servers Database on SSDs Asynchronous processing Application
Assets Analytics

Become price aware Price is a driver of your architecture
r

Automated bidding Automate bids based on on-demand price Run on
spot only when cost e!ective r

Continuous cost optimization Evaluate your reserved instance/on-demand/ spot mix regularly
as app requirements change r

5patterns for performance g DECOUPLED

5patterns for performance g DECOUPLED SCALE OUT

5patterns for performance g DECOUPLED SCALE OUT S OPTIMIZE COST
FOR

FOR v CALIBRATED FOR: CPU performance v CALIBRATED FOR: IO performance

FOR v CALIBRATED FOR: CPU performance v CALIBRATED FOR: IO performance C AUTOMATE

Thank you aws.amazon.com @mza matthew@amazon.com

High Performance Web Applications

High Performance Web Applications

More Decks by Matt Wood

Other Decks in Technology

Featured

Transcript