Aws performance-efficiency-pillar

Performance Efficiency Pillar
AWS Well-Architected Framework
July 2018

© 2018, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Notices
This document is provided for informational purposes only. It represents AWS’s
current product offerings and practices as of the date of issue of this document,
which are subject to change without notice. Customers are responsible for
making their own independent assessment of the information in this document
and any use of AWS’s products or services, each of which is provided “as is”
without warranty of any kind, whether express or implied. This document does
not create any warranties, representations, contractual commitments,
conditions or assurances from AWS, its affiliates, suppliers or licensors. The
responsibilities and liabilities of AWS to its customers are controlled by AWS
agreements, and this document is not part of, nor does it modify, any agreement
between AWS and its customers.

Contents
Introduction 1
Performance Efficiency 1
Design Principles 2
Definition 2
Selection 3
Compute 4
Storage 8
Database 11
Network 19
Review 23
Benchmarking 25
Load Testing 26
Monitoring 27
Active and Passive 28
Phases 29
Trade-Offs 30
Caching 31
Partitioning or Sharding 33
Compression 34
Buffering 35
Conclusion 37
Contributors 37
Further Reading 38
Document Revisions 38

Abstract
The focus of this paper is the Performance Efficiency pillar of the Amazon Web
Services (AWS) Well-Architected Framework. It provides guidance to help
customers apply best practices in the design, delivery, and maintenance of AWS
environments.
This whitepaper covers the best practices for managing production
environments. Designing and managing non-production environments and
processes such as Continuous Integration or Delivery are out of the scope of this
document.

Amazon Web Services – Performance Efficiency Pillar AWS Well-Architected Framework
Page 1
Introduction
The AWS Well-Architected Framework helps you understand the pros and cons
of decisions you make while building systems on AWS. By using the Framework
you will learn architectural best practices for designing and operating reliable,
secure, efficient, and cost-effective systems in the cloud. It provides a way for
you to consistently measure your architectures against best practices and
identify areas for improvement. We believe that having well-architected systems
greatly increases the likelihood of business success.
The framework is based on five pillars:
• Operational Excellence
• Security
• Reliability
• Performance Efficiency
• Cost Optimization
This paper focuses on the performance efficiency pillar and how to apply it to
your solutions. In traditional, on-premises environments, achieving high and
lasting performance can be challenging. By adopting the practices in this paper,
you can build architectures that are efficient and deliver sustained performance
over time.
This paper is intended for those in technology roles, such as chief technology
officers (CTOs), architects, developers, and operations team members. After
reading this paper, you’ll understand AWS best practices and strategies to use
when designing a performant cloud architecture. This paper doesn’t provide
implementation details or architectural patterns. However, it does include
references to appropriate resources for this information.
Performance Efficiency
The performance efficiency pillar focuses on the efficient use of computing
resources to meet requirements and how to maintain that efficiency as demand
changes and technologies evolve. This paper provides in-depth, best-practice
guidance for architecting for performance efficiency on AWS.

Page 2
Design Principles
In the cloud, there are a number of principles that can help you achieve
performance efficiency.
• Democratize advanced technologies: Technologies that are
difficult to implement can become easier to consume by pushing that
knowledge and complexity into the cloud vendor’s domain. Rather than
having your IT team learn how to host and run a new technology, they
can simply consume it as a service. For example, NoSQL databases,
media transcoding, and machine learning are all technologies that
require expertise that is not evenly dispersed across the technical
community. In the cloud, these technologies become services that your
team can consume. This lets your team focus on product development
rather than resource provisioning and management.
• Go global in minutes: Easily deploy your system in multiple AWS
Regions around the world with just a few clicks. This allows you to
provide lower latency and a better experience for your customers at
minimal cost.
• Use serverless architectures: In the cloud, serverless architectures
remove the need for you to run and maintain servers to carry out
traditional compute activities. For example, storage services can act as
static websites, removing the need for web servers, and event services
can host your code for you. This removes the operational burden of
managing these servers and can also lower transactional costs because
these managed services operate at cloud scale.
• Experiment more often: With virtual and automatable resources,
you can quickly carry out comparative testing using different types of
instances, storage, or configurations.
• Mechanical sympathy: Use the technology approach that aligns best
to what you are trying to achieve. For example, consider data access
patterns when you select database or storage approaches.
Definition
Performance efficiency in the cloud is composed of four areas:
• Selection

Page 3
• Review
• Monitoring
• Trade-offs
Take a data-driven approach to building a high-performance architecture.
Gather data on all aspects of the architecture, from the high-level design to the
selection and configuration of resource types. By reviewing your choices on a
regular basis, you ensure that you are taking advantage of the continually
evolving AWS Cloud. Monitoring will ensure that you are aware of any deviance
from expected performance so you can take action on it. Finally, you can make
tradeoffs in your architecture to improve performance, such as using
compression or caching, or relaxing consistency requirements.
Selection
The optimal solution for a particular system will vary based on the kind of
workload you have and often combines multiple approaches. Well-architected
systems use multiple solutions and enable different features to improve
performance.
With AWS, resources are virtualized and are available in a number of different
types and configurations. This makes it easier to find an approach that closely
aligns with your needs. You can also find options that are not easily achievable
with on-premises infrastructure. For example, a managed service such as
Amazon DynamoDB provides a fully managed NoSQL database with single-digit
millisecond latency at any scale.
You can rely on AWS Solutions Architects, AWS Reference Architectures, and
AWS Partners to help you choose an architecture based on what you have
learned. However, you need to perform benchmarking or load testing to
optimize your architecture. After you have identified your architectural
approach, you should use a data-driven process to refine your selection of
resource types and configuration options. You can obtain this data using
benchmarking and load testing. For more information, see the Review section
later in this whitepaper.
Your architecture can combine a number of different architectural approaches
(for example, request-response, event driven; extract, transform, load (ETL); or
pipeline). Your architecture’s implementation will use the AWS services that can

Page 4
optimize your architecture’s performance. In the following sections, we look at
the four main resource types that you should consider: compute, storage,
database, and network.
Compute
The optimal compute solution for a particular system varies based on
application design, usage patterns, and configuration settings. Architectures can
use different compute solutions for various components and enable different
features to improve performance. Selecting the wrong compute solution for an
architecture can lead to lower performance efficiency.
When architecting your use of compute, you should take advantage of elasticity
mechanisms that can ensure that you have sufficient capacity to sustain
performance as demand changes. In AWS, compute is available in three forms:
instances, containers, and functions. Choose which form of compute to use for
each component in your architecture. Instances are generally the default option.
Using containers can improve the utilization of instances. Functions are well
suited to event-driven or highly parallelizable tasks.
Instances
Instances are virtualized servers and, therefore, you can change their
capabilities with the click of a button or an API call. In the cloud, resource
decisions are no longer fixed, so you can experiment with different server types.
Amazon Elastic Compute Cloud (Amazon EC2) virtual server instances come in
different families and sizes. They offer a wide variety of capabilities, including
solid-state drives (SSDs) and graphics processing units (GPUs). When you
launch an EC2 instance, the instance type that you specify determines the
hardware of the host computer used for your instance. Each instance type offers
different compute, memory, and storage capabilities. Instance types are
grouped in instance families based on these capabilities.
When selecting instance families and types also consider the configuration
options that your workload needs:
• Graphics Processing Units (GPU). Using general purpose computing
on GPUs (GPGPU), you can build applications that benefit from the high
degree of parallelism that GPUs provide by leveraging platforms such as

Page 5
CUDA in the development process. Also, if your application requires 3D
rendering or video compression, GPUs enable hardware-accelerated
computation and encoding, making your application more efficient.
• Field Programmable Gate Arrays (FPGA). Using FPGAs, you can
optimize your workloads by having custom hardware-accelerated
execution for your most demanding workloads. You can define your
algorithms by leveraging supported general programming languages
such as C or Go, or hardware-oriented languages such as Verilog or
VHDL.
• Burstable instance families. Burstable instances are designed to
provide moderate baseline performance and the capability to burst to
significantly higher performance when it’s required by your workload.
These instances are intended for workloads that don't use the full CPU
often or consistently, but occasionally need to burst. They are well suited
for general-purpose workloads, such as web servers, developer
environments, and small databases. These instances provide CPU credits
that can be consumed when the instance needs to provide performance.
Credits accumulate when the instance doesn’t need them.
• Advanced computing features. Amazon EC2 gives you access to
advanced computing features such as managing the C-state and P-state
registers and controlling the turbo-boost of processors. Access to co-
processors allows cryptography operations offloading through AES-NI,
or advanced computation through AVX extensions.
You should use data to select the optimal EC2 instance type for your workload,
ensure you have the correct networking and storage options, and consider
operating system settings that can improve the performance for your workload.
Containers
Containers are a method of operating system virtualization that allow you to
run an application and its dependencies in resource-isolated processes.
Amazon EC2 Container Service (Amazon ECS) allows automated execution and
management of containers on a cluster of EC2 instances. You can create services
in Amazon ECS that will be served by containers. By using application Auto
Scaling, you can define automated, metrics-based scaling for your services so
that the number of containers supporting your service grows as the service
needs grow.

Page 6
Amazon ECS leverages Auto Scaling groups, which enable you to scale the
Amazon ECS cluster by adding EC2 instances. This way you ensure that you
always have the appropriate underlying server capacity to host your containers.
Amazon ECS is integrated with Elastic Load Balancing (ELB) to support load
balancing your services in the container fleet dynamically. By creating an
Application Load Balancer instance, your containers will automatically register
themselves to the load balancer when they scale. With these features, you can
host multiple containers serving the same service on a single Amazon EC2
instance, increasing the scalability of your application.
When using containers, you should use data to select the optimal type for your
workload just as you use data to select your EC2 instance types. You should also
consider container configuration options such as memory, CPU, and tenancy
configuration.
Functions
Functions abstract the execution environment from the code you want to
execute. The use of functions allows you to execute code or provide a service
without running or managing an instance.
AWS Lambda lets you run code without provisioning or managing servers or
containers. You just upload your code and AWS Lambda takes care of
everything required to run and scale your code. You can set up your code to
automatically trigger from other AWS services; you can call it directly; or you
can use it with Amazon API Gateway.
Amazon API Gateway is a fully managed service that makes it easy for
developers to create, publish, maintain, monitor, and secure APIs at any scale.
You can create an API that acts as a “front door” to your Lambda function. API
Gateway handles all the tasks involved in accepting and processing up to
hundreds of thousands of concurrent API calls, including traffic management,
authorization and access control, monitoring, and API version management.
To deliver optimal performance with AWS Lambda, you should choose the
amount of memory you want for your function. You are then allocated
proportional CPU power and other resources. For example, choosing 256 MB of
memory allocates approximately twice as much CPU power to your Lambda

Page 7
function as requesting 128 MB of memory. You can also control the amount of
time each function is allowed to run (up to a maximum of 300 seconds).
Elasticity
Elasticity allows you to match the supply of resources you have against demand
for them. Instances, containers, and functions all provide mechanisms for
elasticity either in combination with Auto Scaling or as a feature of the service
itself.
Optimally matching supply to demand delivers the lowest cost for a system, but
you also need to plan for sufficient supply to allow for provisioning time and
individual resource failures. Demand can be fixed or variable, requiring metrics
and automation to ensure that management does not become a burdensome
and disproportionately large cost in itself.
In AWS, you can use a number of different approaches to match supply with
demand. The Cost Optimization Pillar of the AWS Well-Architected
Framework whitepaper describes how to use each of these approaches:
• Demand-based approach
• Buffer-based approach
• Time-based approach
Key AWS Services
The key AWS service for elastic compute solutions is Auto Scaling because you
use it to match the supply of your resources against demand for them.
Instances, containers, and functions all provide mechanisms for elasticity either
in combination with Auto Scaling or as a feature of the service itself.
Resources
Refer to the following resources to learn more about AWS best practices for
compute.
Documentation
• Instances: Instance Types
• Containers: Amazon ECS Container Instances

Page 8
• Functions: Compute Requirements – Lambda Function Configuration
Storage
The optimal storage solution for a particular system varies based on the kind of
access method (block, file, or object) you use, patterns of access (random or
sequential), throughput required, frequency of access (online, offline, archival),
frequency of update (WORM, dynamic), and availability and durability
constraints.
In AWS storage is virtualized, and there are a number of different storage types.
This makes it easier to match your storage methods more closely with your
needs, and it also offers storage options that are not easily achievable with on-
premises infrastructure. For example, Amazon Simple Storage Service (Amazon
S3) is designed for 11 nines of durability. When you use cloud storage, you can
change from using magnetic hard disk drives (HDDs) to solid state drives
(SSDs), and easily move virtual drives from one instance to another in seconds.
Characteristics
When you select a storage solution, you should consider the different
characteristics that you require, such as ability to share, file size, cache size,
latency, throughput, and persistence of data. Then match the requirements you
want to the AWS service that best fits your needs: Amazon S3, Amazon Glacier,
Amazon Elastic Block Store (Amazon EBS), Amazon Elastic File System
(Amazon EFS), or Amazon EC2 instance store.
Performance can be measured by looking at throughput, input/output
operations per second (IOPS), and latency. Understanding the relationship
between those measurements will help you select the most appropriate storage
solution.
Storage Services Latency Throughput Shareable
Block EBS,
EC2 instance
store
Lowest,
consistent
Single Mounted on single instance,
copies via snapshots
File system EFS Low,
consistent
Multiple Many clients
Object S3 Low-latency Web scale Many clients

Page 9
Storage Services Latency Throughput Shareable
Archival Glacier Minutes to
hours
High No
From a latency perspective, if your data is only accessed by one instance then
you should use block storage, such as Amazon EBS with Provisioned IOPS.
Distributed file systems such as Amazon EFS generally have a small latency
overhead for each file operation, so they should be used where multiple
instances need access.
Amazon S3 has features than can reduce latency and increase throughput. You
can use cross-region replication (CRR) to provide lower-latency data access to
different geographic regions.
From a throughput perspective, Amazon EFS supports highly parallelized
workloads (for example, using concurrent operations from multiple threads and
multiple EC2 instances), which enables high levels of aggregate throughput and
operations per second. For Amazon EFS, use a benchmark or load test to select
the appropriate performance mode.
Key AWS Services
The key AWS service for storage is Amazon S3, which provides secure, durable,
highly scalable cloud storage. The following services and features are also
important:
• Amazon EBS provides persistent block storage volumes for use with
EC2 instances.
• Amazon EFS provides simple, scalable file storage for use with EC2
instances.
• Amazon EC2 instance store provides temporary block-level storage
for use with EC2 instances.
• Amazon Glacier provides an archival platform for long-term data
storage and is integrated with Amazon S3 for tiered storage.
Resources
storage.

Page 10
Documentation
• Amazon S3: Request Rate and Performance Considerations
• Amazon Glacier: Amazon Glacier Documentation
• Amazon EFS: Amazon EFS Performance
• Amazon EBS: I/O Characteristics
Videos
• Amazon EBS Design for Performance
Configuration Options
Storage solutions generally have configuration options that allow you to
optimize for the type of workload.
Amazon EBS provides a range of options that allow you to optimize storage
performance and cost for your workload. These options are divided into two
major categories: SSD-backed storage for transactional workloads, such as
databases and boot volumes (performance depends primarily on IOPS), and
HDD-backed storage for throughput-intensive workloads such as MapReduce
and log processing (performance depends primarily on MB/s).
SSD-backed volumes include the highest performance Provisioned IOPS SSD
for latency-sensitive transactional workloads and General Purpose SSD that
balance price and performance for a wide variety of transactional data.
Amazon S3 transfer acceleration enables fast transfer of files over long distances
between your client and your S3 bucket. Transfer acceleration leverages
Amazon CloudFront globally distributed edge locations to route data over an
optimized network path. For a workload in an S3 bucket that has intensive GET
requests you should use Amazon S3 with CloudFront. When uploading large
files, you should use multi-part uploads with multiple parts uploading at once to
help maximize network throughput.
Access Patterns
How you access data will affect how the storage solution performs. Select the
storage solution that aligns best to your access patterns, or consider changing

Page 11
your access patterns to align with the storage solution to maximize
performance.
Creating a RAID 0 (zero) array allows you to achieve a higher level of
performance for a file system than you can provision on a single volume.
Consider using RAID 0 when I/O performance is more important than fault
tolerance. For example, you could use it with a heavily used database where data
replication is already set up separately.
Amazon EBS HDD-backed volumes include Throughput-Optimized HDD for
throughput-intensive sequential workloads and the lowest cost Cold HDD for
less frequently accessed data.
You can further optimize for how you use storage systems by using techniques
covered in the Partitioning or Sharding section of the Trade-Offs topic in this
whitepaper.
Key AWS Services
The key AWS services for storage solutions are Amazon EBS, Amazon S3, and
Amazon EFS which provide block, object, and file system storage solutions.
These services have configuration options that further allow you to optimize
your storage solution. The access patterns your components have should
influence the storage solution you select.
Resources
storage.
Documentation
• Storage
• EBS Storage Types
Database
The optimal database solution for a particular system can vary based on your
requirements for availability, consistency, partition tolerance, latency,
durability, scalability, and query capability. Many systems use different
database solutions for different subsystems and enable different features to
improve performance. Selecting the wrong database solution and features for a
system can lead to lower performance efficiency.

Page 12
Although a workload’s database approach (for example, relational database
management system or RDBMS, NoSQL, etc.) has significant impact on
performance efficiency, it is often an area that is chosen according to
organizational defaults rather than through using a data-driven approach. As
with storage, it’s critical to consider the access patterns of your workload, and
also to consider if other non-database solutions could solve the problem more
efficiently (for example, using an object store such as Amazon S3, a search
engine, or a data warehouse).
In on-premises environments the tendency is to standardize on a common
platform for ease of management. However, in AWS there are services for each
approach. You should consider using different services for different types of
data. You can leverage your own platforms that you deploy on Amazon EC2 or
use fully managed services, allowing you to have the benefits of best-of-breed
solutions without management effort on your side.
Efficiency Factors
There are many factors to consider when selecting your database approach. For
example, your access patterns will help you choose between different types of
database technologies. When selecting your database technology, keep in mind
these four factors:
• Access patterns
• Characteristics
• Configuration options
• Operational effort
Access Patterns
The way that you access data will affect how the database solution performs.
Select the database solution that aligns best to your access patterns, or consider
changing your access patterns to align with the storage solution to maximize
performance.
Optimize how you use database systems based on your access pattern (for
example, indexes, key distribution, partition, or horizontal scaling).

Page 13
Characteristics
Consider the different characteristics you require (for example, availability,
consistency, partition tolerance, latency, durability, scalability, and query
capability) so that you can select the appropriate approach to use (for example,
relational, NoSQL, warehouse).
Configuration Options
Database solutions generally have configuration options that allow you to
optimize for the type of workload. Consider the configuration options for your
selected database approach such as storage optimization, database level
settings, memory, and cache.
Operational Effort
You should consider the effort required to operate your technology choice. The
effort of managing and maintaining a database can vary greatly based on your
technology choice and consumption approach. You should also consider the
effort required to scale these approaches as your requirements grow. For
example, databases that are provided as managed services remove or reduce the
effort of managing database servers and their detailed configuration. To further
optimize your use of database systems, explore the techniques covered in the
Partitioning or Sharding section of the Trade-Offs topic in this whitepaper.
Database Technologies
We commonly see four technology approaches used to address database
requirements:
• Relational Online Transaction Processing (OLTP)
• Non-relational databases (NoSQL)
• Data warehouse and Online Analytical Processing (OLAP)
• Data indexing and searching
Relational Online Transaction Processing (OLTP)
Most traditional applications run on an existing relational database, such as
Oracle, Microsoft SQL Server, MySQL, or PostgreSQL. These database engines
provide high-performance, relational access to datasets. You can make complex
queries that will join and aggregate data from multiple entities. The
transactional model allows complex business requests.

Page 14
You can run these database engines on Amazon EC2 by:
• Installing the software on your own EC2 instances running Microsoft
Windows or Linux
• Using pre-configured Amazon Machine Images (AMIs), provided by
AWS or by third-parties on the AWS Marketplace, with your database
engine installed
There are four distinct tasks for optimizing these databases:
• Optimizing the underlying instance (compute, memory, storage,
network) using the best practices presented in this white paper
• Optimizing the operating system settings such as volume management,
RAID, block sizes, and settings
• Optimizing the database engine configuration and some database-
specific features, such as partitioning
• Optimizing the databases themselves by managing the schemas, indexes,
views and database-related options
Amazon Relational Database Service (Amazon RDS) provides a fully managed
relational database. By leveraging Amazon RDS, you can keep using your
existing database engine (Oracle, Microsoft SQL Server, MySQL, MariaDB,
PostgreSQL) and benefit from an automated environment.
Amazon RDS also proposes a cloud-native database platform called Amazon
Aurora. Aurora is compatible with MySQL or PostgreSQL, and has an execution
and storage environment that is purpose-built for AWS. You will natively
benefit from top-tier performance, with a convenient interface that is
compatible with MySQL or PostgreSQL, and without licensing constraints.
Optimizing an Amazon RDS database comes down to:
• Selecting the appropriate configuration from the catalog of instance
types, and the type of storage required
• Tuning the database engine configuration using the options and
parameter groups

Page 15
• Optimizing the database (schemas, indexes, views, and database-related
options)
All other tasks are managed by the service. Backup, patching, and point-in-time
recovery are automated.
Some database engines provide options for running read replicas. These replicas
allow read-heavy workloads to use multiple servers instead of overloading a
central database.
Amazon RDS provides automated read replicas (not available for Oracle or SQL
Server) to scale out read performance, SSD-backed storage options, and
provisioned IOPS.
Amazon Aurora has a fully redesigned, SSD-based storage layer that scales
automatically as your needs grow, and supports a large number of low-latency
read replicas for increased read performance.
Non-Relational Databases (NoSQL)
New applications can be built without the constraints of managing a central
relational OLTP database system. These systems can leverage scalable
distributed systems such as NoSQL databases. These database platforms scale
out and in by adding nodes to or removing nodes from the database cluster,
making them great candidates for large-scale applications. They also usually
allow less structured data structures to be stored natively. This limits the
burden of schema management and makes them easy to use by application
developers. On the other hand, their scaling nature makes it difficult to run
complex relational queries such as joining tables and entities to have an answer.
This work is usually deferred to the application layer.
You can run NoSQL databases, such as Cassandra, on EC2 instances, design
their storage tier, optimize the operating system, and then run the engine. By
leveraging instance storage and EBS volumes, designing your clusters in
multiple Availability Zones or Regions, and ensuring you have the right model,
you can have high-scale, high-performance platforms for your data needs.
Amazon DynamoDB is a fully managed NoSQL database that provides single-
digit millisecond latency at any scale. For internet-scale information, such as
user profiles, shopping carts, or any other high volume read and write
applications, Amazon DynamoDB can be configured to match your needs. You

Page 16
can specify the exact amount of read and write operations per second that you
expect your table to process. Storage scales up and down automatically based on
your usage of the platform.
Amazon DynamoDB has advanced features that increase the performance and
ease of use of the platform:
• Amazon DynamoDB Accelerator (DAX) provides a read-through/write-
through distributed caching tier in front of the database, supporting the
same API as Amazon DynamoDB, but providing sub-millisecond latency
for entities that are in the cache.
• Amazon DynamoDB also supports throughput Auto Scaling, based on
limits.
Amazon DynamoDB provides indexing features that allow you to efficiently
query on any attribute (column). You can also project attributes (copied from
the table into the index) to improve performance.
Amazon DynamoDB manages table partitioning for you automatically, adding
new partitions if necessary and distributing provisioned throughput capacity
evenly across them. To ensure that your load is evenly distributed across
partitions, you should design for uniform data access across items in your tables
by selecting an appropriate partition key.
Data Warehouse and Online Analytical Processing (OLAP)
OLTP and NoSQL platforms are extremely useful for application deployment,
but often have limitations for large scale analytics. These use cases are better
served by data warehouse platforms that are made for fast access to structured
data. Some advanced technologies are particularly adapted for data
warehousing, such as columnar storage of information, or Massively Parallel
Processing (MPP) environments.
You can host your existing Oracle or Microsoft SQL Server environment
configured for data warehousing workloads. Or you could benefit from
specialized data warehousing platforms such as Pivotal Greenplum.
Alternatively, you could use cloud-native environments, such as Snowflake, by
finding them on the AWS Marketplace.

Page 17
Amazon Redshift is a managed petabyte-scale data warehouse that allows you to
change the number or type of nodes as your performance or capacity needs
change. Use Amazon Redshift when you need SQL operations that will scale.
Amazon Redshift allows you to change the number or type of nodes in your data
warehouse and scale up all the way to a petabyte or more of compressed user
data. Dense compute (DC) nodes allow you to create very high-performance
data warehouses using fast CPUs, large amounts of RAM, and SSDs.
Amazon Redshift will provide the best performance if you specify sort keys,
distribution keys, and column encodings as these can significantly improve
storage, I/O, and query performance.
For increased ease of use, Amazon Redshift also supports a query engine called
Amazon Redshift Spectrum, providing the features of Amazon Redshift on data
directly stored on Amazon S3. With Amazon Redshift Spectrum, you can
request your data stored in Amazon S3 up to exabyte-scale data lake
environments.
Some customers want to run analytics queries on large amounts of less
structured data. Presto is a powerful platform that accepts weakly structured
data. It can query information at large scale from existing storage solutions and
data lakes, such as Amazon S3. Amazon Athena is a fully managed Presto
service that can run queries on your data lakes. Its integration with the data
cataloging features of Amazon Glue and its full serverless model give you strong
query capabilities for your analytics platforms.
Data Indexing and Searching
In many cases, you may need to index, search, and consolidate information at
large scale. This allows your users to access detailed information, and analyze
this information following multiple angles. Technologies such as search engines
are particularly efficient for these use cases. Use search engines when you want
search or report with low latency and high throughput.
Apache Lucene is a search engine platform that is widely used worldwide
because it has been integrated in multiple platforms, such as Elasticsearch.
Elasticsearch provides an easy-to-use platform that can automatically discover
and index documents at a really large scale. Elasticsearch is also at the center of
a complete ecosystem that supports reporting with Kibana and document or log

Page 18
aggregation with LogStash and other providers. Kibana provides a simple
solution for creating dashboards and analysis on indexed data.
You can deploy Elasticsearch on EC2 instances, leveraging EBS volumes, in
multiple Availability Zones for high-availability and high performance.
Elasticsearch supports scale-out, which is increasing capacity by adding new
nodes to the cluster.
Amazon Elasticsearch Service (Amazon ES) is a managed service in the AWS
Cloud that makes it simple and cost-effective to set up, manage, and scale an
Elasticsearch cluster for your cloud application.
Amazon ES offers powerful scaling features on disk size or instance
configuration (CPU, memory) and a model you can scale-out by adding nodes.
As your data or query volume changes, you can scale your search domain's
resources up or down as needed. You can control scaling if you know that you
need more capacity for bulk uploads or are expecting a surge in search traffic.
Key AWS Services
The key AWS services for database solutions are Amazon RDS, Amazon
DynamoDB, and Amazon RedShift, which provide relational, NoSQL, and data
warehouse solutions. These services have configuration options that further
allow you to optimize your storage solution. The access patterns of your
components should influence the storage solution you select.
Resources
databases.
Documentation
• Cloud Databases with AWS
• Amazon Aurora best practices
• Amazon Redshift performance
• Amazon Athena top 10 performance tips
• Amazon Redshift Spectrum best practices

Page 19
• Amazon DynamoDB best practices
• Amazon DynamoDB Accelerator
Network
The optimal network solution for a particular system will vary based on latency,
throughput requirements, and so on. Physical constraints such as user or on-
premises resources will drive location options, which can be offset using edge
techniques or resource placement.
In AWS, networking is virtualized and is available in a number of different types
and configurations. This makes it easier to match your networking methods
more closely with your needs. AWS offers product features (for example,
enhanced networking instance types, Amazon EBS optimized instances,
Amazon S3 transfer acceleration, and dynamic CloudFront) to optimize network
traffic. AWS also offers networking features (for example, Amazon Route 53
latency routing, Amazon Virtual Private Cloud (Amazon VPC) endpoints, and
AWS Direct Connect) to reduce network distance or jitter.
Location
The AWS Cloud infrastructure is built around Regions and Availability Zones. A
Region is a physical location in the world having multiple Availability Zones.
Availability Zones consist of one or more discrete data centers, each with
redundant power, networking, and connectivity, housed in separate facilities.
These Availability Zones offer you the ability to operate production applications
and databases that are more highly available, fault tolerant, and scalable than
would be possible from a single data center.
Choose the appropriate Region or Regions for your deployment based on some
key elements:
• Where your users are located: Choosing a Region close to your
application’s users ensures lower latency when they use the application.
• Where your data is located: For data-heavy applications, the major
bottleneck in latency is when data is transferred to the computing part of
the application. Application code should execute as close to the data as
possible.

Page 20
• Other constraints: Take into account constraints such as security and
compliance.
Placement Groups
Amazon EC2 provides placement groups for networking. A placement group is a
logical grouping of instances within a single Availability Zone. Using placement
groups with supported instance types enables applications to participate in a
low-latency, 20-gigabits-per-second (Gbps) network. Placement groups are
recommended for applications that benefit from low network latency, high
network throughput, or both. Using placement groups has the benefit of
lowering jitter in network communications.
Edge Locations
Latency-sensitive services are delivered at the edge using a global network of
edge locations. These edge locations commonly provide services such as Content
Delivery Network (CDN) and Domain Name System (DNS). By having these
services at the edge, they can respond with low latency to requests for content or
DNS resolution. These services can also provide geographic services such as Geo
Targeting of content (providing different content based on end users’ location),
or latency-based routing to direct end users to the nearest Region (minimum
latency).
CloudFront is a global CDN that can be used to accelerate both static content
such as images, scripts, and videos, as well as dynamic content such as APIs or
web applications. It relies on a global network of edge locations that will cache
the content and provide high-performance network connectivity to our users.
CloudFront also accelerates many other features such as content uploading and
dynamic applications, making it a performance addition to all applications
serving traffic over the internet.
Route 53 is a highly available and scalable cloud DNS web service. It’s designed
to give developers and businesses an extremely reliable and cost-effective way to
route end users to internet applications by translating names, like
www.example.com, into numeric IP addresses, like 192.168.2.1, that computers
use to connect to each other. Route 53 is fully compliant with IPv6 as well.
You should use edge services to reduce latency and to enable caching of content.
You need to ensure that you have configured cache control correctly for both
DNS and HTTP/HTTPS to gain the most benefit from these approaches.

Page 21
Product Features
In cloud computing networking is critical, and therefore services commonly
offer features to optimize network performance. You should consider product
features such as EC2 instance network capability, enhanced networking
instance types, Amazon EBS-optimized instances, Amazon S3 transfer
acceleration, and dynamic CloudFront to optimize network traffic.
Amazon S3 content acceleration is a feature that lets external users benefit from
the networking optimizations of CloudFront to upload data to Amazon S3. This
makes it easy to transfer large amounts of data from remote locations that don’t
have dedicated connectivity to the AWS Cloud.
EC2 instances can also leverage enhanced networking. Enhanced networking
uses single root I/O virtualization (SR-IOV) to provide high-performance
networking capabilities on supported instance types. SR-IOV is a method of
device virtualization that provides higher I/O performance and lower CPU
utilization than traditional virtualized network interfaces. Enhanced networking
provides higher bandwidth, higher packet per second (PPS) performance, and
consistently lower inter-instance latencies.
Amazon Elastic Network Adapters (ENA) provide further optimization by
delivering 20 Gbps of network capacity for your instances within a single
placement group.
Amazon EBS–optimized instances use an optimized configuration stack and
provide additional, dedicated capacity for Amazon EBS I/O. This optimization
provides the best performance for your EBS volumes by minimizing contention
between Amazon EBS I/O and other traffic from your instance.
Networking Features
When architecting your solution in the cloud you should consider networking
features to reduce network distance or jitter.
Latency-based routing (LBR) for Route 53 helps you improve your application’s
performance for a global audience. LBR works by routing your customers to the
AWS endpoint (for EC2 instances, Elastic IP addresses, or ELB load balancers)
that provides the fastest experience based on actual performance measurements
of the different AWS Regions where your application is running.

Page 22
AWS Direct Connect provides dedicated connectivity to the AWS environment,
from 50 Mbps up to 10 Gbps. This gives you managed and controlled latency
and provisioned bandwidth so your applications can connect easily and in a
performant way to other environments. Using one of the AWS Direct Connect
partners, you can have end-to-end connectivity from multiple environments,
thus providing an extended network with consistent performance.
Amazon VPC endpoints provide reliable connectivity to AWS services (for
example, Amazon S3) without requiring an internet gateway or a Network
Address Translation (NAT) instance. Using VPC endpoints, the data between
your VPC and another AWS service is transferred within the Amazon network,
helping protect your instances from internet traffic.
When implementing a scale-out architecture where you want to use multiple
instances for service content, you can leverage load balancers inside your
Amazon VPC. AWS provides multiple models for your applications in the ELB
service. Application Load Balancer is best suited for load balancing of HTTP and
HTTPS traffic and provides advanced request routing targeted at the delivery of
modern application architectures, including microservices and containers.
Network Load Balancer is best suited for load balancing of TCP traffic where
extreme performance is required. It is capable of handling millions of requests
per second while maintaining ultra-low latencies, and it is also optimized to
handle sudden and volatile traffic patterns.
Key AWS Services
The key AWS service for networking solutions is Route 53, which provides
latency-based routing. In addition, the use of Amazon VPC endpoints and AWS
Direct Connect can reduce network distance or jitter.
Resources
networking.
Documentation
• Networking Products with AWS

Page 23
Review
When you first architect your solution, you have a fixed set of options to choose
from. Over time, new technologies and approaches become available that can
improve the performance of your architecture. In the AWS Cloud, it’s much
easier to experiment with new features and services because your infrastructure
is code.
Using AWS, you can take advantage of our continual innovation, which is driven
by customer need. We reduce prices and release new Regions, edge locations,
services, and features regularly. Any of these new releases could positively
improve the performance efficiency of your architecture.
After you have identified your architectural approach, you should use
benchmarking and load testing data to drive your selection of resource types
and configuration options.
To adopt a data-driven approach to architecture you should implement a
performance review process that considerers the following:
• Infrastructure as code: Define your infrastructure as code using
approaches such as AWS CloudFormation templates. The use of
templates allows you to place your infrastructure into source control
alongside your application code and configurations. This enables you to
apply the same practices you use to develop software to your
infrastructure so you can iterate rapidly.
• Deployment pipeline: Use a continuous integration/continuous
deployment (CI/CD) pipeline (for example, source code repository, build
systems, deployment, and testing automation) to deploy your
infrastructure. This enables you to deploy in a repeatable, consistent,
and low-cost fashion as you iterate.
• Well-defined metrics: Set up your metrics and monitoring to capture
key performance indicators (KPIs). We recommend that you use both
technical and business metrics. For website or mobile apps key metrics
are capturing time to first byte or rendering. Other generally applicable
metrics include thread count, garbage collection rate, and wait states.
Business metrics, such as the aggregate cumulative cost per request, can
alert you to ways to drive down costs. Carefully consider how you plan to

Page 24
interpret metrics. For example, you could choose the maximum or 99th
percentile instead of the average.
• Performance test automatically: As part of your deployment
process, automatically trigger performance tests after the quicker
running tests have passed successfully. The automation should create a
new environment, set up initial conditions such as test data, and then
execute a series of benchmarks and load tests. Results from these tests
should be tied back to the build so you can track performance changes
over time. For long running tests, you can make this part of the pipeline
asynchronous from the rest of the build. Alternatively, you could execute
performance tests overnight using Amazon EC2 Spot Instances.
• Load generation: You should create a series of test scripts that
replicate synthetic or prerecorded user journeys. These scripts should be
idempotent and not coupled, and you might need to include “pre-
warming” scripts to yield valid results. As much as possible, you want
your test scripts to replicate the behavior of usage in production. You
can use software or software-as-a-service (SaaS) solutions to generate
the load. Consider using AWS Marketplace solutions and Spot Instances.
They can be cost-effective ways to generate the load.
• Performance visibility: Key metrics should be visible to your team,
especially metrics against each build version. This allows you to see any
significant positive or negative trend over time. You should also display
metrics on the number of errors or exceptions to make sure you are
testing a working system.
• Visualization: Use visualization techniques that make it clear where
performance issues, hot spots, wait states, or low utilization is occurring.
Overlay performance metrics over architecture diagrams-- call graphs or
code can help identify issues more quickly.
This performance review process can be implemented as a simple extension of
your existing deployment pipeline and then evolved over time as your testing
requirements become more sophisticated. For future architectures, you should
be able generalize your approach and reuse the same process and artifacts.
When architectures perform badly this is normally because a performance
review process has not been put into place or is broken. If your architecture is
performing badly, putting this performance review process in place will allow

Page 25
you to apply Deming’s plan-do-check-act (PDCA) cycle to drive iterative
improvement.
Benchmarking
Benchmarking uses synthetic tests to provide you with data on how components
perform. In this section, we discuss how to use benchmarking to drive
improvements in your workloads. We don’t cover the use of benchmarking to
compare different vendor’s products or implementations.
Benchmarking is generally quicker to set up than load testing and is used when
you want to evaluate the technology for a particular component. Benchmarking
is often used at the start of a new project, when you don’t yet have a whole
solution that you could load test.
For benchmarking, you should follow the performance review process.
However, your deployment pipeline will just consist of the benchmark tests
themselves. You can either build your own custom benchmark tests, or you can
use an industry standard test, such as TPC-DS (to benchmark your data
warehousing workloads). Industry benchmarks are helpful when you are
comparing different environments. Custom benchmarks are useful for targeting
specific types of operations that you expect to make in your architecture.
With benchmarking, it is generally more important to pre-warm your test
environment to ensure valid results. You should run the same benchmark
multiple times to be sure you’ve captured any variance over time.
Since benchmarks are generally faster to run than load tests, they can be used
earlier in the deployment pipeline to provide faster feedback on performance
deviations to your team. When you evaluate a significant change in a component
or service, a benchmark can be a quick way to see if you can justify the effort to
make the change based on the benchmarked difference. Benchmarking should
be used in conjunction with load testing because load testing will tell you how
your whole workload will perform in production.
Key AWS Services
The key AWS services supporting benchmarking are AWS CodeDeploy and AWS
CloudFormation. Use these services to automate the testing of your
infrastructure in a repeatable fashion.

Page 26
Resources
benchmarking.
Videos
• Performance Channel
• Performance Benchmarking on AWS
Documentation
• Amazon S3 Performance Optimization
• Amazon EBS Volume Performance
• AWS CodeDeploy
• AWS CloudFormation
Load Testing
Load testing uses your actual workload so you can see how your whole solution
performs in a production environment. Load tests should be done using
synthetic or sanitized versions of production data (remove sensitive or
identifying information). Use replayed or pre-programmed user journeys
through your application at scale that exercise your entire architecture. As part
of your delivery pipeline you can automatically carry out load tests and compare
against pre-defined KPIs and thresholds to ensure that you continue to get the
performance you require.
Amazon CloudWatch can collect metrics across the resources in your
architecture. You can also collect and publish custom metrics to surface
business or derived metrics. Use CloudWatch to set alarms that indicate when
thresholds are breached to signal that a test is outside of expected performance.
Using AWS services, you can run production-scale environments to test your
architecture aggressively. Since you only pay for the test environment when it is
needed, you can carry out full-scale testing at a fraction of the cost of using an
on-premises environment. You should take advantage of the AWS Cloud to test
your workload to see where it fails to scale or scales in a non-linear way. You can

Page 27
use Spot Instances to generate loads at low cost and discover bottlenecks before
they are experienced in production.
When you write critical user stories for your architecture, you should include
performance requirements such as specifying how quickly each critical story
should execute. For these critical stories, you should implement additional
scripted user journeys to ensure you have visibility into how these stories
perform against your requirement.
Where load tests take considerable time to execute, you should parallelize them
using multiple copies of your test environment. Your costs will be similar, but
your testing time will be reduced. (It costs the same to run one EC2 instance for
100 hours as it does to run 100 instances for one hour.) You can also lower the
costs of these tests by using Spot Instances and selecting Regions that have
lower costs than the Regions you use for production.
The location of your load test clients should reflect the geographic spread of
your end users.
Key AWS Services
The key AWS service supporting load testing is CloudWatch, which allows you
to collect metrics on how your whole architecture is performing during the load
test. With CloudWatch you can also create custom and business metrics.
Resources
Refer to the following resources to learn more about AWS best practices for load
testing.
Documentation
• Load Testing CloudFront
Monitoring
After you have implemented your architecture you will need to monitor its
performance so that you can remediate any issues before your customers are
aware of them. Monitoring metrics should be used to raise alarms when
thresholds are breached. The alarm can trigger automated action to work
around any badly performing components.

Page 28
CloudWatch provides the ability to monitor and send notification alarms. You
can use automation to work around performance issues by triggering actions
through Amazon Kinesis, Amazon Simple Queue Service (Amazon SQS), and
AWS Lambda.
Active and Passive
Monitoring solutions fall into two types: active monitoring (AM) and passive
monitoring (PM). AM and PM complement each other to give you a full view of
how your workload is performing.
Active monitoring simulates user activity in scripted user journeys across
critical paths in your product. AM should be continuously performed in order to
test the performance and availability of a workload. AM complements PM by
being continuous, lightweight, and predictable. It can be run across all
environments (especially pre-production environments) to identify problems or
performance issues before they affect end users.
Passive monitoring is commonly used with web-based workloads. PM
collects performance metrics from the browser (non-web-based workloads can
use a similar approach). You can collect metrics across all users (or a subset of
users), geographies, browsers, and device types. You should use PM to
understand these issues:
• User experience performance: PM provides you with metrics on
what your users are experiencing, which gives you a continuous view
into how production is working, as well as a view into the impact of
changes over time.
• Geographic performance variability: If a workload has a global
footprint and users access the application from all around the world,
using PM can enable you to spot a performance problem affecting users
in a specific geography.
• The impact of API use: Modern workloads use internal APIs and
third-party APIs. PM provides the visibility into the use of APIs so you
can identify performance bottlenecks that originate not only from
internal APIs but also from third-party API providers.

Page 29
Phases
Monitoring at AWS consists of five distinct phases, which are explained in more
detail in the Reliability Pillar of the AWS Well-Architected Framework
whitepaper:
1. Generation – scope of monitoring, metrics, and thresholds
2. Aggregation – creating a complete view from multiple source
3. Real-time processing and alarming – recognizing and responding
4. Storage – data management and retention policies
5. Analytics – dashboards, reporting, and insights
CloudWatch is a monitoring service for AWS Cloud resources and the workloads
that run on AWS. You can use CloudWatch to collect and track metrics, collect
and monitor log files, and set alarms. CloudWatch can monitor AWS resources
such as EC2 instances and RDS DB instances, as well as custom metrics
generated by your applications and services, and any log files your applications
generate. You can use CloudWatch to gain system-wide visibility into resource
utilization, application performance, and operational health. You can use these
insights to react quickly and keep your application running smoothly.
CloudWatch dashboards enable you to create reusable graphs of AWS resources
and custom metrics so you can monitor operational status and identify issues at
a glance.
Ensuring that you do not see too many false positives or are overwhelmed with
data is key to an effective monitoring solution. Automated triggers avoid human
error and can reduce the time it takes to fix problems. Plan for game days,
where simulations are conducted in the production environment, to test your
alarm solution and ensure that it correctly recognizes issues.
Key AWS Services
The key AWS service that supports monitoring is Amazon CloudWatch, which
allows easy creation of alarms that can trigger scaling actions. The following
services and features are also important:
• Amazon S3 acts as the storage layer and allows for lifecycle policies
and data management.

Page 30
• Amazon EMR can analyze metrics to help you monitor performance.
Resources
monitoring to promote performance efficiency.
Videos
• AWS re:Invent 2015 | (DVO315) Log, Monitor and Analyze your IT with
Amazon CloudWatch
• AWS re:Invent 2015 | (BDT312) Application Monitoring: Why Data
Context Is Critical
Documentation
• CloudWatch Documentation
Trade-Offs
When you architect solutions, think about trade-offs so you can ensure an
optimal approach. Depending on your situation you could trade consistency,
durability, and space versus time or latency, to deliver higher performance.
Using AWS, you can go global in minutes and deploy resources in multiple
locations across the globe to be closer to your end users. You can also
dynamically add read-only replicas to information stores, such as database
systems to reduce the load on the primary database. AWS also offers caching
solutions such as Amazon ElastiCache, which provides an in-memory data store
or cache, and CloudFront, which caches copies of your static content closer to
end users.
The following sections detail some of the trade-offs you can make and how you
can implement them.
Technique Applies To Use Gains
Caching Read-heavy Space (Memory) Time
Partitioning or
Sharding
Write-heavy Size & Complexity Time
Compression Large data Time Space

Page 31
Technique Applies To Use Gains
Buffering Many requests Space & Time Efficiency
Caching
Most workloads rely on a dependent component such as a service or database
that offers a source of truth or a consolidated view of data. Generally, these
architecture components are harder to scale and represent a significant
proportion of the cost of the workload. You can improve performance efficiency
by using caching to trade off against freshness or memory used. These
techniques generally update asynchronously or periodically. The trade-off is
that your data isn’t always fresh and, therefore, not always consistent with the
source of truth.
Application Level
You can make this trade-off at a code level by using application-level caches or
memorization. When requests are cached, execution time is reduced. This
provides a way to scale horizontally through the caching layer and reduces load
on your most heavily used components.
In-memory and distributed caches are used in two main cases: coordinating
transient data and state between distinct servers, such as user sessions in an
application, or protecting databases from read-heavy workloads by serving the
most requested elements directly from memory.
Platforms such as Redis, Memcached, or Varnish can be deployed on Amazon
EC2 and will provide robust caching engines for your applications. Design and
scale of these platforms are usually based on the memory of the instance and by
designing appropriate key management in the applications when you want to
activate a cluster model (for example, for Redis clustering, ElastiCache-
consistent hashing).
Amazon ElastiCache is a web service that makes it easy to deploy, operate, and
scale an in-memory data store or cache in the cloud. ElastiCache with
Memcached supports sharding to scale in-memory cache with multiple nodes.
ElastiCache for Redis includes clustering, with multiple shards forming a single
in-memory key-value store that is terabytes in size, plus read replicas per shard
for increased data access performance.

Page 32
ElastiCache is a fully managed service that supports scaling the environment,
automated failover, patching, and backup.
Database Level
Database replicas enhance performance databases by replicating all changes to
the master databases to read replicas (not available for Oracle or SQL Server).
This replication makes it possible to scale out beyond the capacity constraints of
a single database for read-heavy database workloads.
Amazon RDS provides read replicas as a fully managed service. You can create
one or more replicas of a given source database and serve high-volume
application read traffic from multiple copies of your data, thereby increasing
aggregate read throughput. You should also add additional indexes to the read
replica, where the database engine supports it. For example, you could add
more indexes to the MySQL read replica. For latency-sensitive workloads you
should use the Multi-AZ feature to specify which Availability Zones the read
replica should be in to reduce cross-Availability Zone traffic.
Geographic Level
Another example of caching is using a CDN, which is a good way to reduce
latency for clients. CDNs should be used to store static content and to accelerate
dynamic content. You should consider using a CDN for your APIs. Even
dynamic content can benefit through the use of network optimization methods.
CloudFront can be used to deliver your entire website, including dynamic,
static, streaming, and interactive content using a global network of edge
locations. Requests for your content are automatically routed to the nearest
edge location, so content is delivered with the best possible performance.
CloudFront is optimized to work with other AWS services, like Amazon S3,
Amazon EC2, ELB, and Route 53. CloudFront also works seamlessly with any
non-AWS origin server, which stores the original, definitive versions of your
files.
Key AWS Services
The key AWS services for caching solutions are ElastiCache, which provides a
general-purpose application cache, and CloudFront, which allows you to cache
information closer to your users.

Page 33
Resources
caching.
Documentation
• Best Practices for Implementing Amazon ElastiCache
• AWS CloudFormation Best Practices
Video
• Turbocharge your apps with Amazon ElastiCache
Partitioning or Sharding
When using technologies, such as relational databases, that require a single
instance due to consistency constraints, you can only scale vertically (by using
higher specification instances and storage features). When you hit the limits of
vertical scaling, you can use a different approach called data partitioning or
sharding. With this model, data is split across multiple database schemas, each
running in its own autonomous primary DB instance.
Amazon RDS removes the operational overhead of running multiple instances,
but sharding will still introduce complexity to the application. The application’s
data access layer will need to be modified to have awareness of how data is split
so that it can direct queries to the right instance. (You can use a proxy or routing
mechanism to remove caching code from the application, or implement it in a
data access layer.) In addition, any schema changes will have to be performed
across multiple database schemas, so it is worth investing some effort to
automate this process.
NoSQL database engines will typically perform data partitioning and replication
to scale both the reads and the writes in a horizontal fashion. They do this
transparently without the need of having the data partitioning logic
implemented in the data access layer of your application. Amazon DynamoDB
in particular manages table partitioning for you automatically, adding new
partitions as your table grows in size or as read- and write-provisioned capacity
changes.

Page 34
Partitioning or sharding provides a way to scale write-heavy workloads, but
requires that data is evenly distributed and evenly accessed across all partitions
or shards. It can introduce complexity in relational database solutions, while
NoSQL solutions generally trade consistency to deliver this.
Key AWS Services
The key AWS service for partitioning or sharding is Amazon DynamoDB, which
manages table partitioning for you automatically.
Resources
partitioning and sharding.
Documentation
• Best Practices for DynamoDB
Video
• AWS re:Invent 2015 | (DAT401) Amazon DynamoDB Deep Dive
Compression
Compressing data trades computing time against space and can greatly reduce
storage and networking requirements. Compression can apply to file systems,
data files, and web resources such as stylesheets and images, but also to
dynamic responses such as APIs.
CloudFront supports compression at the edge. The source system can serve
resources in a standard fashion, and the CDN will automatically compress the
resources if and only if the web clients can support it.
When transferring large quantities of information into or out of the cloud you
should consider non-network based solutions. AWS Snowball is a petabyte-scale
data transport solution that uses secure appliances to transfer large amounts of
data into and out of the AWS Cloud. Using AWS Snowball addresses common
challenges with large-scale data transfers, including high network costs, long
transfer times, and security concerns.

Page 35
Amazon Redshift uses compression with columnar data storage. Instead of
storing data as a series of rows, Amazon Redshift organizes the data by column.
Columnar data stores can be compressed much more than row-based data
stores because similar data is stored sequentially on disk. Amazon Redshift
employs multiple compression techniques and can often achieve significant
compression relative to traditional relational data stores. When loading data
into an empty table, Amazon Redshift automatically samples your data and
selects the most appropriate compression scheme.
Key AWS Services
The key AWS service for compression is CloudFront, which supports
compression at the edge.
Resources
compression.
Documentation
• Amazon CloudFront: Serving Compressed Files
• Amazon RedShift: Columnar Storage
• AWS Snowball: What Is AWS Snowball?
Buffering
Buffering uses a queue to accept messages (units of work) from producers. For
resiliency, the queue should use durable storage. A buffer is a mechanism to
ensure that applications can communicate with each other when they are
running at different rates over time. Messages can then be read by consumers,
which allows the messages to run at the rate that meets the consumers’ business
requirements. By using a buffer, you can decouple the throughput rate of
producers from that of consumers. You don’t have to worry about producers
having to deal with data durability and backpressure (where producers slow
down because their consumer is running slowly).
When you have a workload that generates significant write load that doesn’t
need to be processed immediately, you can use a buffer to smooth out demands
on consumers.

Page 36
On AWS, you can choose from multiple services to implement a buffering
approach. Amazon SQS provides a queue that allows a single consumer to read
individual messages. Amazon Kinesis provides a stream that allows many
consumers to read the same messages.
An Amazon SQS queue is a reliable, scalable, and fully managed repository for
messages that are awaiting processing. Amazon SQS enables applications to
quickly and reliably queue messages that one component in the application
generates to be consumed by another component. For this approach, a queue is
created that producers post messages to, and which resides at a well-known
address.
To reduce cost, producers post messages using Amazon SQS batch API actions.
Consumers read messages from the well-known queue using a fixed-sized Auto
Scaling group of EC2 instances to cope with instance failures. Long polling lets
you retrieve messages from your Amazon SQS queue as soon as they become
available. Use long polling to reduce the cost of retrieving messages. Amazon
SQS uses the message visibility feature to hide messages that have been read.
Message visibility reduces the risk that you might process the same message
twice, though you should be aware that by default Amazon SQS at-least-once
delivery means that you can receive a message more than once.
You can use Amazon SQS first in/first out (FIFO) queues if you need exactly-
once processing. Message visibility allows for messages that have not been
processed to reappear (for example, in the case of an instance failure). For long-
running tasks you can extend the visibility time-out. Your application will need
to delete messages after they have been processed.
An alternative approach is to use Amazon Kinesis to provide buffering. It differs
from Amazon SQS in that it allows multiple consumers to read the same
message at any one time. However, a single message will be read using the
Kinesis Client Library, or AWS Lambda, and will be delivered to one-and-only-
one consumer for an Application, which is a name provided by the consumer
when it connects to the Stream. Different Applications can all consume the
same messages concurrently, providing a “Publish and Subscribe” model.
When architecting with a buffer keep in mind two key considerations. First,
what is the acceptable delay between producing the work and consuming the
work? Second, how do you plan to handle duplicate requests for work?

Page 37
To optimize the speed with which work items are consumed by more
consumers, you can use Spot Instances. Using Spot Instances, you can bid on
spare Amazon EC2 computing capacity. To handle duplicate messages with
Amazon SQS we recommend using Amazon SQS first in-first-out (FIFO)
queues, which provide exactly-once processing.
Key AWS Services
The key AWS service for buffering is Amazon SQS, which allows you to decouple
producers from consumers using a queue.
Resources
Refer to the following resource to learn more about AWS best practices for
buffering.
Documentation
• Best Practices for Amazon SQS
Conclusion
Achieving and maintaining performance efficiency requires a data-driven
approach. You should actively consider access patterns and trade-offs that will
allow you to optimize for higher performance. Using a review process based on
benchmarks and load tests will allow you to select the appropriate resource
types and configurations. Treating your infrastructure as code will enable you to
rapidly and safely evolve your architecture, while you use data to make fact-
based decisions about your architecture. Putting in place a combination of
active and passive monitoring will ensure that the performance of your
architecture does not degrade over time.
AWS strives to help you build architectures that have performance efficiency
while delivering business value. To make your architectures truly perform, you
should use the tools and techniques discussed in this paper.
Contributors
The following individuals and organizations contributed to this document:

Page 38
• Philip Fitzsimons, Sr Manager Well-Architected, Amazon Web Services
• Julien Lépine, Specialist SA Manager, Amazon Web Services
• Ronnen Slasky, Solutions Architect, Amazon Web Services
Further Reading
For additional help, please consult the following sources:
• AWS Well-Architected Framework
Document Revisions
Date Description
July 2018 Minor update for grammatical issues
November 2017 Refreshed the whitepaper to reflect changes in AWS
November 2016 First publication

Aws performance-efficiency-pillar

More Related Content

What's hot (20)

Similar to Aws performance-efficiency-pillar (20)

More from Darnette A (10)

Recently uploaded (20)

Aws performance-efficiency-pillar