Performance-Tweaks für Java-Serverless-Anwendungen (Java Forum Nord 2023)

JAVA FORUM NORD 2023 – PERFORMANCE-TWEAKS FÜR JAVA-SERVERLESS-ANWENDUNGEN
© 2023, Amazon Web Services, Inc. or its affiliates. All rights reserved.
Performance-Tweaks
für Java-Serverless-
Anwendungen
Dennis Kieselhorst
Principal Solutions Architect
Amazon Web Services

View Slide

What is Serverless?
No infrastructure provisioning,
no management
Automatic scaling
Pay for value Highly available and secure

View Slide

Let‘s start with a non-serverless architecture…
3

View Slide

…and transform it to a first serverless architecture
4

View Slide

© 2023, Amazon Web Services, Inc. or its affiliates. All rights reserved. 5
1
2 Request
Request
Initialize application
Running application
3 Request …
• Initialize the environment once
• Handle multiple request with the same instance (concurrently)
• Instance keeps running after request processing
• Scaling based on metrics or manually
Java in a container or a VM
Initialization

View Slide

1
Initialization Execution
Execution
2
Execution
3
Initialization
4
Execution
Initialize Application
Initialize Application
Java with Serverless functions
• Initialize execution environment on first invoke
• An execution environment handles a single request at a time
• If an execution environment is busy with a request – another environment
is started – simpler programming model
• Scales to zero if there is no request – Managed auto scaling

View Slide

What are AWS Lambda functions?
7
Handler function
• Function executed on invocation
• Processes incoming event
Event
• Invocation data sent to function
• JSON event payload differs by source (API, State change, Queue)
Context
• Additional information from the service
• Example: Request ID

View Slide

Tune the serverless function's resources…
AWS Lambda exposes only a memory control, with the
% of CPU core and network capacity allocated to a function
proportionally.
Is your code CPU, network or memory-bound?
If so, it could be cheaper to increase memory.
> Memory, > vCPU, > Network
10240 MB
128 MB

View Slide

…with Power Tuning
9
https://github.com/alexcasalboni/aws-lambda-power-tuning

View Slide

Leverage multithreading
I/O bound workloads will likely see gains by multithreading.
< 1 vCPU: CPU bound workloads won't see gains
> 1 vCPU: you can leverage multiple vCPUs
10240 MB
128 MB
6 vCPU
1 vCPU
1769 MB

View Slide

Custom AWS silicon with 64-bit Arm cores
Targeted optimizations to reduce costs
Rapidly innovate, build, and iterate
Range of instance types
Same AWS Nitro building blocks
Best performance per watt of energy use in
Amazon EC2
AWS Graviton2 &
Graviton3
X86 vs. ARM – AWS Graviton processor

View Slide

The lifecycle of your AWS Lambda function
Time
Load and
initialize
handler
JVM start
Run
handler
code
Compile
Package
Deploy
Download
code
Build and deploy Initialize Handler invocation

View Slide

Time
Load and
initialize
handler
JVM start
Run
handler
code
Compile
Package
Deploy
Cold start
Download
code
Warm start
Build and deploy Handler invocation
Initialize

View Slide

Provisioned Concurrency
• Managed solution to avoid „cold starts“
• You set provisioned concurrency for individual functions
• No code changes required
• Integrated with AWS Auto Scaling
• Additional costs
14

View Slide

Time
Load and
initialize
handler
JVM start
Run
handler
code
Compile
Package
Deploy
Download
code
Boosted host CPU access
Initialize
up to 10 seconds

View Slide

Time
Eliminate or move activities to earlier phases
Load and
initialize
handler
JVM start
Run
handler
code
Compile
Package
Deploy
Download
code
Initialize
up to 10 seconds
Avoid: normal Lambda charges incur

View Slide

Monitor and tune the Garbage Collection
Monitor the Garbage
Collection logs, try a fixed
heap size specific to your
application.
Example shows 100 calls
with heap size 400 MB:
• 5% call duration
reduction
• Only minor garbage
collection activities

View Slide

Tiered Compilation Setting
Set JVM tiered compilation level to 1 and use the C1
compiler.
JAVA_TOOL_OPTIONS=
-XX:+TieredCompilation
-XX:TieredStopAtLevel=1
C1 compiler quickly produces native code.
No profiling overhead, but no benefits from more
aggressive C2 compilation.
related blogpost
Enabled by default in AWS
Lambda Java 17 runtime!

View Slide

Dependencies
• Trade-off between feature richness & performance
• Examples
▪ jackson-jr for serializiation
▪ SimpleLogger for logging
▪ Spring FunctionHandler instead of RestController (directly handle
API-Gateway events instead of HTTP)
• Be careful with dependency injection: Annotation scanning adds
more class loading.
20

View Slide

AWS SDK and Common Runtime Client
• As of today there are two version of the AWS SDK for Java (AWS
SDK for Java v1 and AWS SDK for Java v2 ).
• We recommend to use the latest version AWS SDK for Java v2
because it has been designed to improve the performance and
usability.
• The AWS Common Runtime (CRT) HTTP client is a new HTTP client
you can use with the AWS SDK for Java v2. The CRT-based HTTP
client is an asynchronous, non-blocking HTTP client built on top of
the Java bindings of the AWS Common Runtime.
21

View Slide

AWS Lambda
SnapStart
Up to 10x faster startup
performance
NEW

View Slide

AWS Lambda SnapStart overview
Publish
Init
Encrypted snapshot
stored
Tiered low-
latency cache
Invoke
microVM snapshot technology

View Slide

Invoke
Resume
Invoke
Resume
Invoke
Resume
Invoke
Resume
Invoke
Resume
The AWS Lambda SnapStart lifecycle
Publish
Init
Encrypted snapshot
stored
Tiered low-
latency cache
Invoke
Invoke
Resume Invoke
[…]

View Slide

Time
JVM start
Run
handler
code
Compile
Package
Deploy
Download
code
Build and
Deploy
Handler
Invocation
Initialize
Publish Invoke
Load and
initialize
handler
Resize
Heap

View Slide

Considerations
Network connections Ephemeral data Uniqueness

View Slide

Runtime Hooks
• Runtime hooks are available as part of the open-source
Coordinated Restore at Checkpoint (CRaC) project.
• Include dependency (e.g. Maven)

io.github.crac
org-crac
0.1.3

27

View Slide

CRaC interface for post- and pre-hooks examples
- beforeCheckpoint
- priming
- class loading
- dependency injection
- afterRestore
- unique ID for execution
environment
- re-establish connections

View Slide

Time
JVM start
Run
handler
code
Compile
Package
Deploy
Download
code
Build and
Deploy
Handler
Invocation
Initialize
Publish Invoke
After
Restore()
Before
Snapshot()
Load and
initialize
handler
Resize
Heap

View Slide

Time
Run
handler
code
Compile
Package
Deploy Configure
API clients
Reflection and
Class Loading
Dependency
Injection
JIT
Compilation
Build and
Deploy
Handler
Invocation
Initialize
After
Restore()
Before
Snapshot()
JVM start
Download
code
Load and
initialize
handler
Resize
Heap
Publish Invoke

View Slide

SnapStart benefits
• Start your functions from pre-initialized snapshots
• No additional cost
• Minimal to no code changes
• Customization via Runtime Hooks
31
Without SnapStart
With SnapStart
6.725
- 80 %
1.15
Full Duration
Results will vary based on your application

View Slide

Database connection pooling
• Database connection pools (e.g. HikariCP) initialize multiple database
connections during startup
• Since an AWS Lambda environment handles a single request at a time
you only need 1 connection per environment
• Ensure to stay within database connection pool limits
• Solution: RDS Proxy or Reserved Concurrency (Limit connections)
…
Container

View Slide

Powertools for AWS Lambda
33
J A V A L I B R A R Y T O S I M P L I F Y T R A C I N G , S T R U C T U R E D L O G G I N G A N D C U S T O M M E T R I C S
• Facilitate best practice adoption
• Increase developer velocity
• Optional modules:
• Parameter handling
• Idempotency
• SQS Large Message and Batching Optiongs
• Serialization, Validation and custom resource support

View Slide

Compile your application ahead-of-time (AOT)
• GraalVM creates a native image of your application at build-time.
• Frameworks like Quarkus, Micronaut, and Spring Native integrate
GraalVM and add further benefits like build-time dependency
injection.
• Some JVM features not supported.
• Application startup time reduced of more than 90%

View Slide

GraalVM native images on AWS Lambda
• Create a native binary and deploy as AWS Lambda custom runtime
• Sub-second application startup and low memory footprint
• Multiple frameworks provide built-in support (Micronaut, Quarkus, Spring)
35
bootstrap

View Slide

Future
• The Java ecosystem is adapting and introduced several new
projects to foster innovation.
• Project Leyden will address startup time, memory footprint and
peak performance by introducing a concept of static images to the
Java Platform, and to the JDK.
• The changes to the release cycle mean that innovation will be
delivered faster.
37

View Slide

Deep Dive: Java on AWS Lambda Workshop
38
https://catalog.workshops.aws/java-on-aws-lambda

View Slide

Thank you!

View Slide

Performance-Tweaks für Java-Serverless-Anwendungen (Java Forum Nord 2023)

Performance-Tweaks für Java-Serverless-Anwendungen (Java Forum Nord 2023)

Dennis Kieselhorst

More Decks by Dennis Kieselhorst

Other Decks in Programming

Featured

Transcript