OpenChain - The Ramifications of ISO/IEC 5230 and ISO/IEC 18974 for Legal Pro...
IPT High Performance Reactive Java BGOUG 2016
1. Trayan Iliev
CEO of IPT – Intellectual
Products & Technologies
tiliev@iproduct.org
http://iproduct.org
BGOUG Conference
June 3-5 ‘16, Borovetz
High Performance
Reactive Programing
with Java 8
2. 2
IPT – Intellectual Products &
Technologies: IT Education Evolved
Since 2003 we provide trainings and share skills &
knowledge in Java SE/ EE/ Web/ JS/ Angular / React:
Java EE/Web, JSF 2, Portlets, Portals: Liferay, GateIn
Reactive IoT with Reactor / RxJava / RxJS
Angular 2 + Ionic 2+ TypeScript web/mobile client MV*
frameworks
REST HATEOAS – Distributed Hypermedia APIs
Domain Driven Design, Reactive Microservices,
CQRS and Event Sourcing
Oracle® & Java™ are (registered) trademarks of Oracle and/or its affiliates. Liferay® is
registered trademark of Liferay, Inc. Other names may be trademarks of their respective owners.
3. 3
High Performnce Reactive JAVA
Reactive programming. Reactor & Proactor design
patterns. Reactive Streams (java.util.concurrent.Flow)
High performance non-blocking asynchronous apps
on JVM using Reactor project & RxJava
Disruptor (RingBuffer), Flux & Mono, Processors
End-to-end reactive web applications and services:
Reactor IO (REST, WebSocket) + RxJS + Angular 2
Demo - reactive hot event streams processing on
Raspberry Pi 2 (ARM v7) based robot IPTPI.
RxJava (not Zen only :) coans for self assessment
4. 4
The Internet of Things has the potential to change the
world, just as the Internet did. Maybe even more so.
Nearly 50 petabytes of data are captured and created
by human beings
People have limited time, attention and accuracy
Capturing data about things in the real world in real time
Track and count everything, reduce waste, loss & cost.
Know when things need replacing, repairing or recalling
— Kevin Ashton, 'That 'Internet of Things' Thing', RFID Journal,
2009
Internet of Things (IoT)
5. 5
Robots: The Most Intelligent Things
CC BY 2.0, Source:
https://www.flickr.com/photos/wilgengebroed/8249565455/
Radar, GPS, lidar for navigation and obstacle
avoidance ( 2007 DARPA Urban Challenge )
6. 6
Tracking Complexity
We need tools to cope with all that complexity inherent in
robotics and IoT domains.
Simple solutions are needed – cope with problems through
divide and concur on different levels of abstraction:
Domain Driven Design (DDD) – back to basics:
domain objects, data and logic.
Described by Eric Evans in his book:
Domain Driven Design: Tackling Complexity in the Heart of
Software, 2004
7. 7
Be Reactive: What It Really Means?
My Favourite Definition of
Reactive Streams :)
https://youtu.be/qybUFnY7Y8w
8. 8
Imperative and Reactive
We live in a Connected Universe
... there is hypothesis that all
the things in the Universe are
intimately connected, and you
can not change a bit without
changing all.
Action – Reaction principle is
the essence of how Universe
behaves.
9. 9
Imperative and Reactive
Reactive Programming: using static or dynamic data
flows and propagation of change
Example: a := b + c
Functional Programming: evaluation of mathematical
functions,
− Avoids changing-state and mutable data, declarative
programming
− Side effects free => much easier to understand and
predict the program behavior.
Example: books.stream().filter(book -> book.getYear() > 2010)
.forEach( System.out::println )
10. 10
Functional Reactive (FRP)
According to Connal Elliot's (ground-breaking paper @
Conference on Functional Programming, 1997), FRP is:
(a) Denotative
(b) Temporally continuous
13. 13
Reactive Streams Spec.
Reactive Streams – provides standard for
asynchronous stream processing with non-blocking
back pressure.
Minimal set of interfaces, methods and protocols for
asynchronous data streams
April 30, 2015: has been released version 1.0.0 of
Reactive Streams for the JVM (Java API,
Specification, TCK and implementation examples)
Java 9: java.util.concurrent.Flow
14. 14
Reactive Streams Spec.
Publisher – provider of potentially unbounded number
of sequenced elements, according to Subscriber(s)
demand.
Publisher.subscribe(Subscriber) => onSubscribe onNext*
(onError | onComplete)?
Subscriber – calls Subscription.request(long) to
receive notifications
Subscription – one-to-one Subscriber ↔ Publisher,
request data and cancel demand (allow cleanup).
Processor = Subscriber + Publisher
15. 15
FRP = Async Data Streams
FRP is asynchronous data-flow programming using the
building blocks of functional programming (e.g. map,
reduce, filter) and explicitly modeling time
Used for GUIs, robotics, and music. Example (RxJava):
Observable.from(
new String[]{"Reactive", "Extensions", "Java"})
.take(2).map(s -> s + " : on " + new Date())
.subscribe(s -> System.out.println(s));
Result:
Reactive : on Wed Jun 17 21:54:02 GMT+02:00 2015
Extensions : on Wed Jun 17 21:54:02 GMT+02:00 2015
16. 16
Performance is about 2 things (Martin Thompson –
http://www.infoq.com/articles/low-latency-vp ):
– Throughput – units per second, and
– Latency – response time
Real-time – time constraint from input to response
regardless of system load.
Hard real-time system if this constraint is not honored then
a total system failure can occur.
Soft real-time system – low latency response with little
deviation in response time
100 nano-seconds to 100 milli-seconds. [Peter Lawrey]
What About High Performance?
17. 17
Mechanical Sympathy – hardware (CPU, cache, memory,
IO, Network), operating system, language implementation
platform (e.g. JVM), and application level code are working
in harmony to minimize the time needed for event (request,
message) processing => 10% / 90% principle
Throughput vs. latency – bus vs. car traveling
Throughput ~ System Capacity / Latency
Achieving low latency may mean additional work done
by system => lowered System Capacity and Throughput
Horizontal scalability is valuable for high throughput. For
low latency, you need simplicity – critical path.
Throughput vs. Latency
18. 18
JVMs are often faster than custom C++ code because of the
holistic optimizations that they can apply across an application
[Andy Piper].
Developers can take advantage of hardware guarantees through
a detailed understanding of:
– Java Memory Model & mapping to underlying hardware
– low latency software system hardware (CPU, cache, memory,
IO, Network)
– avoiding lock-contention and garbage collection
– Compre-And-Swap – CAS (java.util.concurrent.atomic)
– lock-free, wait-free techniques – using standard libraries (e.g.
the LMAX Disruptor)
High Performance Java
19. 19
CPU Cache – False Sharing
Core 2 Core NCore 1 ...
Registers
Execution Units
L1 Cache A | | B |
L2 Cache A | | B |
L3 Cache A | | B |
DRAM Memory
A | | B |
Registers
Execution Units
L1 Cache A | | B |
L2 Cache A | | B |
20. 20
Low garbage by reusing existing objects + infrequent GC
when application not busy – can improve app 2 - 5x
JVM generational GC startegy – ideal for objects living very
shortly (garbage collected next minor sweep) or be immortal
Non-blocking, lockless coding or CAS
Critical data structures – direct memory access using
DirectByteBuffers or Unsafe => predictable memory layout
and cache misses avoidance
Busy waiting – giving the CPU to OS kernel slows program
2-5x => avoid context switches
Amortize the effect of expensive IO - blocking
Low Latency: Things to Remember
21. 21
Parallel tasks can increase your throughput by increasing
system capacity – it is GOOD!
But comes together with concurrent access to shared
resources => you have to provide mutual exclusion (MutEx)
by parallel threads when changing the resources' state (read
only access can be shared by multiple threads)
Mutual exclusion can be achieved in several ways:
– synchronized – hardwired in HotSpot JVM, optimized in J^6
– ReentrantLock, ReadWriteLock, StampedLock →
java.util.concurrent.locks.*
– Optimistic Locking → tryLock(), CAS
Parallelism & Concurrency
22. 22
Simple problem: incrementing a long value 500 000 000 times.
9 implementations:
‒ SynchronousCounter – while (counter++ < 500000000){}
‒ SingleThreadSynchronizedCounter – 1T using synchronized
‒ TwoThreadsSynchronizedCounter – 2T using synchronized
‒ SingleThreadCASCounter – 1T using AtomicLong
‒ TwoThreadsCASCounter – 2T using AtomicLong
‒ TwoThreadsCASCounterLongAdder – 1T using LongAdder
‒ SingleThreadVolatileCounter – 1T, memory barrier (volatile)
‒ TwoThreadsVolatileCounter – 2T, memory barrier (volatile)
Comparing Concurrent Impl.
23. 23
Test results (on my laptop - quad core Intel i7@2.2GHz):
− SynchronousCounter – 190ms
− SingleThreadSynchronizedCounter – 15000 ms
− TwoThreadsSynchronizedCounter – 21000 ms
− SingleThreadCASCounter – 4100 ms
− TwoThreadsCASCounter – 12000 ms
− TwoThreadsCASCounterLongAdder – 12800 ms
− SingleThreadVolatileCounter – 4100 ms
− TwoThreadsVolatileCounter – 20000 ms
Comparing Concurrent Impl.
24. 24
For more complete micro-benchmarking of
different Mutex implementations see:
http://blog.takipi.com/java-8-stampedlocks-vs-
readwritelocks-and-synchronized/
http://www.slideshare.net/haimyadid/java-8-
stamped-lock
Comparing Concurrent Impl.
25. 25
Non-blocking (synchronous) implementation is 2 orders of
magnitude better then synchronized
We should try to avoid blocking and especially contended
blocking if want to achieve low latency
If blocking is a must we have to prefer CAS and optimistic
concurrency over blocking (but have in mind it always
depends on concurrent problem at hand and how much
contention do we experience – test early, test often,
microbenchmarks are unreliable and highly platform dependent
– test real application with typical load patterns)
The real question is: HOW is is possible to build concurrency
without blocking?
Mutex Comparison => Conclusions
26. 26
Message Driven – asynchronous message-passing allows
to establish a boundary between components that ensures
loose coupling, isolation, location transparency, and
provides the means to delegate errors as messages
[Reactive Manifesto].
The main idea is to separate concurrent producer and
consumer workers by using message queues.
Message queues can be unbounded or bounded (limited
max number of messages)
Unbounded message queues can present memory
allocation problem in case the producers outrun the
consumers for a long period → OutOfMemoryError
Scalable, Massively Concurrent
27. 27
Queues typically use either linked-lists or arrays for the
underlying storage of elements. Linked lists are not
„mechanically sympathetic” – there is no predictable
caching “stride” (should be less than 2048 bytes in each
direction).
Bounded queues often experience write contention on
head, tail, and size variables. Even if head and tail
separated using CAS, they usually are in the same cache-
line.
Queues produce much garbage.
Typical queues conflate a number of different concerns –
producer and consumer synchronization and data storage
Queues Disadvantages
[http://lmax-exchange.github.com/disruptor/files/Disruptor-1.0.pdf]
28. 28
LMAX Disruptor design pattern separates different
concerns in a “mechanically sympathetic” way:
- Storage of items being exchanged
- Producer coordination – claiming the next sequence
- Consumers coordination – notified new item is available
Single Writer principle is employed when writing data in
the Ring Buffer from single producer thread only (no
contention),
When multiple producers → CAS
Memory pre-allocated – predictable stride, no garbage
LMAX Disruptor (RingBuffer)
[http://lmax-exchange.github.com/disruptor/files/Disruptor-1.0.pdf]
29. 29
LMAX Disruptor (RingBuffer) High Performance
[http://lmax-exchange.github.com/disruptor/files/Disruptor-
1.0.pdf]
Source: LMAX Disruptor github wiki - https://raw.githubusercontent.com/wiki/LMAX-
Exchange/disruptor/images/Models.png
LMAX-Exchange Disruptor License @ GitHub: Apache License Version 2.0, January 2004 -
http://www.apache.org/licenses/
30. 30
LMAX Disruptor (RingBuffer) High Performance
[http://lmax-exchange.github.com/disruptor/files/Disruptor-
1.0.pdf]
Source: LMAX Disruptor @ GitHub - https://github.com/LMAX-
Exchange/disruptor/blob/master/docs/Disruptor.docx
LMAX-Exchange Disruptor License @ GitHub: Apache License Version 2.0, January 2004 -
http://www.apache.org/licenses/
31. 31
Project Reactor
Reactor project allows building high-performance (low
latency high throughput) non-blocking asynchronous
applications on JVM.
Reactor is designed to be extraordinarily fast and can
sustain throughput rates on order of 10's of millions of
operations per second.
Reactor has powerful API for declaring data
transformations and functional composition.
Makes use of the concept of Mechanical Sympathy
built on top of Disruptor / RingBuffer.
32. 32
Project Reactor
Pre-allocation at startup-time
Message-passing structures are bounded
Using Reactive and Event-Driven Architecture patterns
=> non-blocking end-to-end flows, replies
Implement Reactive Streams Specification – efficient
bounded structures requesting no more than capacity
Applies above features to IPC and provides non-
blocking IO drivers that are flow-control aware
Expose a Functional API – organize their code in a
side-effect free way, which helps you determine you are
thread-safe and fault-tolerant
45. 45
LeJaRo: Lego®
Java Robot
Modular – 3 motors (with encoders) – one driving each
track, and third for robot clamp.
Three sensors: touch sensor (obstacle avoidance), light
color sensor (follow line), IR sensor (remote).
LeJaRo is programmed in Java using LeJOS library.
More information about LeJaRo:
http://robolearn.org/lejaro/
Programming examples available @GitHub:
https://github.com/iproduct/course-social-robotics/tre
e/master/motors_demo
LEGO® is a registered trademark of LEGO® Group. Programs of IPT are not
affiliated, sponsored or endorsed by LEGO® Education or LEGO® Group.
47. 47
Let's See Some Real Code
Reactive Websocket Demo available @GitHub:
https://github.com/iproduct/ipt-angular2-
reactive-websocket-demo
Reactive Robotics Demo available @GitHub:
https://github.com/iproduct/jprime-demo
59. 59
Takeaways: Why Go Reactive?
Benefits using Reactive Programming + DDD:
DDD helps to manage complexity in IoT and Robotics -
many subsystems = sub-domains
Reactive Streams (Fluxes, Monos) = uni-directional data
flows, CQRS, event sourcing, microservices
Reactive Streams can be non-blocking and highly
efficient, or can utilize blocking if needed
Naturally implement state management patterns like
Redux, allow time travel, replay and data analytics
Clear, declarative data transforms that scale (Map-
Reduce, BigData, PaaS)
60. 60
Takeaways: Why Maybe Not?
Cons using Reactive Programming + DDD:
DDD requires additional efforts to clearly separate
different (sub) domains – DSL translators, factories...
Reactive Streams utilize functional composition and
require entirely different mindset then imperative – feels
like learning foreign language
Pure functions and Redux provide much benefits,
but there's always temptation to “do it the old way” :)
Tool support for functional programming in Java is still
not perfect (in Eclipse at least :)
61. 61
Resources: RxMarbles & Rx Coans
RxMarbles:
http://rxmarbles.com/
RxJava Koans – Let's try to solve them at:
https://github.com/mutexkid/rxjava-koans
RxJS Koans – for those who prefer JavaScript :)
https://github.com/Reactive-Extensions/RxJSKoans
62. 62
Interested in Reactive?
IPT Reactive Java and Angular 2 + Typescript
courses: http://iproduct.org
More information about robots @RoboLearn:
http://robolearn.org/
TuxCon – 9-10 July, Plovdiv:
http://tuxcon.mobi/#schedule