Advanced Introduction to Java Multi-Threading - Full (chok)

Full Version

Advanced Introduction
to Java Multi-
Threading
LAU Chok Sheak, July 2012
chok.lau@citi.com

Programmer/Analyst
Trading Services Team,
Citi Equities Technology

Public

Synopsis
What is an "advanced introduction to Java multi-
threading"?
Advanced Talks about concepts not easily obvious
to Java developers.
Introduction Broad overview of what can be done
with Java threads.
Multi-threading Talks about interaction among different
threads.
This presentation is suitable for Java developers at all levels
(beginner to expert).

Public

Synopsis (cont'd)
Why is an advanced introduction needed?

1. Goes underneath the hood of JVM to see
what kind of multi-threaded code is
correct and what isn’t.

2. Explores utilities and libraries available to
the Standard JDK developer.

Public

Structure of this Presentation
This presentation has two parts:

Part I. Java memory model

Part II. Concurrent utilities

Public

Part I. Java Memory Model
Question:
Do I need to know what is the Java Memory Model?

Answer:
If you only write single-threaded programs => No.
If you write multi-threaded programs => Yes.

Public

What is the Java Memory Model (JMM)?
• Some good, standard definitions can be easily found
online.
• The JMM is defined in JSR-133.
To rephrase in my own way, the JMM is:
A. Machine-level – A set of rules to define how CPU
caches should be affected by Java code.
B. Source-level – A set of rules to define whether a
multi-threaded Java program will work correctly or
not, given three areas of consideration:
1. Visibility
2. Ordering
3. Atomicity

Public

JMM – Sample Code 1: Non-volatile
conflicts
Consider the following Java code snippet:

Initialization Thread 1 Thread 2
(both not (not synchronized) (not synchronized)
volatile)

int n = 0; n = 1; if (v == true)
boolean v = false; v = true; System.out.println("n=" + n);

Question: Is it possible for Thread 2 to print “n=0”?

Public

JMM – Sample Code 1: Non-volatile
conflicts (cont'd)
(both not (not synchronized) (not synchronized)
volatile)

int n = 0; n = 1; if (v == true)
boolean v = false; v = true; System.out.println("n=" + n);

Answer: Yes! (may print “n=0”)
Why? v=true might come before n=1 in compiler reorderings.
Also n=1 might not become visible in Thread 2, so Thread 2
will load n=0.
Fix: If you make v as volatile, then Thread 2 can only print
“n=1”. n does not need to be volatile. n cannot be reordered
to come after v=true when v is volatile.

Public

JMM – Sample Code 2: Synchronized
block

int x = 0; synchronized (obj) { synchronized (obj) {
int y = 0; x = x + 1; if (x != y)
final Object obj = y = x; System.out.println("x!=y");
new Object(); } }

Execution: Threads 1 and 2 run their own code repeatedly in
their own infinite loops. Both x and y are not volatile.
Synchronization is by obj, not by the instance object "this".

Question: Is it possible for Thread 2 to print "x!=y"?

Public

JMM – Sample Code 2: Synchronized
block (cont'd)

int x = 0; synchronized (obj) { synchronized (obj) {
int y = 0; x = x + 1; if (x != y)
final Object obj = y = x; System.out.println("x!=y");
new Object(); } }

Answer: No! (will never print “x!=y”)
Why? The values of both x and y get synchronized between
Threads 1 and 2 by implicit memory barriers across the
synchronized block boundaries. All values will be
synchronized even if they are not volatile. The
synchronization object does not have to be the owning class
instance object "this". It can be any object.

Public

JMM – Sample Code 3: Double-
checked locking
public class MyClass { Question: Is it possible
private MyClass() {} to create two or more
private static MyClass instance; instances of MyClass
public static MyClass using getInstance()
getInstance() { when this code is run
if (instance == null) in multi-threaded
synchronized (MyClass.class) { mode?
if (instance == null)
instance = new
Hint: Does "instance"
MyClass();
need to be volatile in
}
return instance;
order for this to work
}
correctly?
}
Public

checked locking (2)
Answer: No, but the code is buggy.
Why? "instance" must be volatile, otherwise
the invocation of the constructor could be
reordered out of the synchronized block
and a different thread might get "instance"
in an uninitialized state even though it is
not possible to create two instances here.

Public

checked locking (3)
Some people claim that this code:
instance = new MyClass();
Can be compiled into this code (which is correct):
A. instance = create new uninitialized object
B. invoke constructor on instance
Which means that between points A and B, instance is in
an uninitialized state, and another thread may obtain a
reference to it. Therefore this code is only correct in a
single-threaded context, but not in a multithreaded
context. This is possible when instance is not volatile
and thus can be reordered!

Public

JMM and Double-Checked Locking
There are two cases to consider to see whether you
should use double-checked locking or not:
Case 1: Instance initialization is cheap.
If it is very quick to create a new instance and the
instance does not take up much memory, then
you should simply use:
public static final MyClass instance = new
MyClass();
There is no need to save on creating this instance if
it does not require much resources.

Public

JMM and Double-Checked Locking (2)
Case 2: Instance initialization is expensive.
Some instances might be expensive to create, then
lazy init should be used. Examples of expensive
init:
• Runs a heavyweight algorithm to compute
certain values.
• Consumes much memory (maybe over a few
thousand bytes)
• Creates a new thread (could be a new
java.lang.Timer, which is a thread)

Public

Then you can use the below code to do • Guaranteed to be visible
lazy initialization: to all threads that use
that class.
public class SingletonExample { • Inner class is not loaded
private static final class until some thread
InnerSingleton { references one of its
fields or methods.
private static final SingletonExample
s= • Maybe looks a bit
cleaner than using
new SingletonExample();
double-checked locking.
}
public static SingletonExample
• Uses classloader
synchronization, which
getInstance() {
could be a bit more
return InnerSingleton.s; efficient.
}
}
Public

 Joshua Bloch recommends using volatile
with local variable to cache the field value.
 http://java.sun.com/developer/technicalAr
ticles/Interviews/bloch_effective_08_qa.ht
ml)
 Feel free to use this idiom if you like.

Public

Part I. Java Memory Model

A. Machine-Level Memory Model

B. Source-Level Memory Model

Public

Machine-Level Memory Model
• The JMM is a noteworthy achievement.
• Given that every CPU architecture
(x86, ia64, ARM, PPC etc.) is
different, you can really appreciate that
Java has defined the JMM to standardize
the execution of all Java code when
executed on any platform.
• Similar thing is done for the C# CLR.
(maybe for my next presentation!)

Public

Machine-Level Memory Model (2)
Given a particular variable x (whether static,
instance, volatile or anything) in Java,
when you access (read or write) the
variable, you could be referring to one or
all of these physical entities:
• CPU register (super-fast)
• CPU cache (very fast)
• Main memory (slow)

Public

Machine-Level Memory Model (3)
In order to synchronize accesses to these
machine-level entities, you have to make
sure that the following software observe
the same JMM rules:
1. Java compiler and its optimizations
(including instruction reorderings)
2. JVM and its emitted memory barrier
instructions (including instruction
reorderings)

Public

Memory Barriers
What is a Memory Barrier?
It is a machine-level instruction to ensure that any
previous loads/stores/monitors in any thread will be
made visible to the current load/store/monitor in a
particular thread.
They are meaningful only when you talk about multi-
threaded programs.
Only JVMs (not Java bytecode compilers) need to worry
about Memory Barriers. However, Java developers
should also be conscious of it because it might come in
handy.

Public

Types of Java Memory Operations
All Java memory operations are one of six
types:

1. Normal Load (non-volatile)
2. Normal Store (non-volatile)
3. Volatile Load
4. Volatile Store
5. Monitor Enter
6. Monitor Exit
Public

Types of Java Memory Barriers
(All theoretically possible)
LoadLoad LoadStore LoadEnter LoadExit

StoreLoad StoreStore StoreEnter StoreExit
EnterLoad EnterStore EnterEnter EnterExit

ExitLoad ExitStore ExitEnter ExitExit

So basically, all combinations of regex
"((Load)|(Store)|(Enter)|(Exit)){2}" - 24 possibilities.

Memory Barriers are used to enforce strict ordering or reading from
main memory or flushing caches to main memory.

Any more details would be beyond the scope of this presentation, but
you may feel free to read up further on your own.

Public

How to Use Memory Barriers in Java?
Java Developers have indirect control of the
usage of memory barriers by using these
Java keywords:
1. volatile
2. synchronized
3. final

Many details in JSR-133. We will cover
some, but not all.

Public

JMM – 1. Visibility
Rule 1: Two accesses (reads or writes) to the same variable
are conflicting if any one of them is a write.
Rule 2: For two conflicting volatile accesses, or non-volatile
accesses across a synchronized block boundary, they are
guaranteed to happen one after the other, and the later
access will never read/overwrite a stale value.
Rule 3: A volatile read can never read a value that has never
been written to it, i.e. cannot create a value out of thin air.
Rule 4: You will always read the same value from a final
variable after the constructor setting the final variable exits.
(If you try to read it before the constructor exits, you might
read a different value.)

Public

JMM – 2. Ordering
When we say "instructions reordering", we are
referring to both the Java bytecode compiler and
JVM hotspot compiler.

Rule 1: Instructions can never be reordered around
volatile variables.
Rule 2: Instructions can never be reordered around
synchronization boundaries, except when we are
moving instructions into the critical section.
Rule 3: Accesses to final fields cannot be reordered.

Public

JMM – 3. Atomicity
Rule 1: Word-tearing: Bytes, booleans, shorts,
integers and floats will never be overwritten by
adjacent writes (whether in array or not).
Rule 2: Reads and writes of volatile doubles and
longs are always atomic, but no guarantee of
atomicity for non-volatile doubles and longs.
Rule 3: Reads and writes of references (whether
32-bit or 64-bit) are always atomic (regardless of
whether they are volatile or not).

Public

JMM – Some Technical Terms
Reading JSR-133 is not exactly fun and it is not easy to
understand. Here are two technical terms being used in JSR-
133 whose meanings might not be obvious:
Happens-before: Means that within a single thread, if
instruction A happens-before B, then B is guaranteed to see
the value of all variables modified by A after the
modification.
Synchronizes-with: Means that if instruction A in Thread T1
happens-before B in Thread T2, then B is guaranteed to see
the value of all variables modified by A after the
modification.
Hence, happens-before and synchronizes-with essentially
mean the same thing, except happens-before is for single
threads, while synchronizes-with is for multiple threads.

Public

Applying JMM in Your Code
 Check variables accessed by multiple threads. Do
they need to be volatile/synchronized? (not
always yes)
 Note that reads can be reordered and also need
to be synchronized! (or made volatile) Don’t think
that only writes need to be synchronized.
 Write multi-threaded, high-volume, stressful test
cases and run them on machines with multiple
CPUs.
 Prefer using volatiles and CAS over
synchronization as volatiles and CAS are very
fast.

Public

End of Part I. Java Memory Model
You might have some questions at this
point, but not to worry, you will be able to
find all the answers online.



Public

Part II. Concurrent Utilities
Brief overview of what is available under
java.util.concurrent.*.

• Lock objects
• Executors
• Concurrent collections
• Atomic variables
• Other stuff

Public

Concurrent Utilities
1. Lock Objects
Package: java.util.concurrent.locks
ReentrantLock
• Almost the same as the synchronized keyword with wait(), notify() and
notifyAll(), but gives more flexibility.
• "Reentrant" means that when you have already acquired the lock, you
can acquire the same lock again without deadlocking yourself.
• "Reentrant" also means that if two or more threads execute the same
piece of code at the same time, there won’t be any conflicts (this
definition does not apply here).
• The synchronized keyword and the ReentrantLock uses two completely
different sets of code and has nothing in common.
• The synchronized keyword seems to perform better with uncontended
locks, but the ReentrantLock seems to perform better with contended
locks. (no guarantees here)
• Main use of ReentrantLock – can try locking with timeout.

Public

1. Lock Objects (2)
ReentrantReadWriteLock
• Encapsulates two different types of locks within
one lock – a read lock and a write lock.
• Usually only useful when you have a large
number of readers and small number of writers,
and user operations outweigh synchronization
overhead.
• Usually only useful when you have multiple
CPUs.

Public

2. Executors
• Abstraction of the execution of a large number of
tasks immediately or in the future.
• Basically only three types of Executors in the
JDK:
o ThreadPoolExecutor
o ScheduledThreadPoolExecutor
o ForkJoinPool (new in Java 7)

Public

2. Executors (2)
• You might find that ThreadPoolExecutor is the
only executor you will ever need to use.
• Option of either fixed or variable sized thread
pools.
• I have never found variable-sized thread pools
useful, so I only need fixed-sized thread pools
(thread creation is a bit expensive).
• Make sure never to use an unbounded thread
pool (unlimited number of threads).

Public

3. Concurrent Collections
ConcurrentHashMap
• Perhaps the most commonly-used concurrent
collection that exists.
• Similar to a synchronized map, but allows more
concurrency by splitting the table into lock
sections.
• Performance is comparable to Cliff Click’s
NonBlockingHashMap, which is a lock-free
hashmap that outperforms ConcurrentHashMap
only when you have a large number of
processors.

Public

3. Concurrent Collections (2)
ArrayBlockingQueue
• Fixed-length array-backed blocking queue.

LinkedBlockingQueue
• Variable-length linked-list-backed blocking
queue.
• Allows option to specify max possible queue
size so that it does not get filled up
indefinitely.

Public

3. Concurrent Collections (3)
ArrayBlockingQueue vs LinkedBlockingQueue
• Both use very basic wait/notifyAll implementation
with no spin retries.
• Advantage of using linked list: 1) saves memory
when queue is empty.
• Advantage of using array: 1) less memory allocation
and GC, 2) less pointer dereferencing (faster), 3)
uses less memory when queue is full.
• Hence most of the time I will choose to use an
ArrayBlockingQueue, not a LinkedBlockingQueue.

Public

3. More Concurrent Collections
• Some more concurrent collections that I have
never used before.
• Provided here only as an overview.
• Source code and javadoc readily available online.

ConcurrentLinkedDeque DelayQueue
ConcurrentLinkedQueue ForkJoinPool (new in Java 7)
ConcurrentSkipListMap LinkedBlockingDeque
ConcurrentSkipListSet LinkedTransferQueue
CopyOnWriteArrayList PriorityBlockingQueue
CopyOnWriteArraySet SynchronousQueue

Public

The LMAX Disruptor Blocking Queue
• The LMAX Disruptor is perhaps the fastest blocking queue
implementation in the planet.
• Boasts a speed of 100K events per second (10 nanoseconds per event)
• Main difference from ArrayBlockingQueue is that the Disruptor uses spin
retries. Spin retries prevent the threads from doing lock, park, and
release immediately when there is no work to do, hence saving CPU
cycles.
• Spin retries are slower when your input is slow (does more work).
• Disruptor is faster than the ArrayBlockingQueue only at super high
throughputs (100K per second). At slower throughputs, the Disruptor is
the same speed or slower than the ArrayBlockingQueue.
• I implemented an AtomicBlockingQueue that does something similar, and
is slightly faster than the Disruptor, but difference is hardly noticeable.
• Conclusion: Your application can hardly catch up with these speeds, so
using the ArrayBlockingQueue and LinkedBlockingQueue should be your
best options!

Public

4. Atomic Variables
Package: Those that I have never
java.util.concurrent.atomic used:
AtomicIntegerArray
AtomicLongArray
Utilities to provide atomic
updates to variables AtomicMarkableReference
without the overhead of AtomicReference
synchronization. AtomicReferenceArray
AtomicStampedReference

Most commonly used:
AtomicBoolean
AtomicInteger
AtomicLong

Public

What is CAS?
• Atomic variables work using a principle/algorithm called
Compare-And-Set (CAS).
• CAS is supported at the machine instruction level. Hence it
is very fast.
The basic idea is like this:
1. Get the expected value (usually current value)
2. Get the update value (usually expected value + delta)
3. Check that the variable currently has the expected value
4. Set the variable to the update value
5. Check that the variable now has the update value
6. If true, return success. Else retry from 1.

Public

CAS Algorithm
The getAndIncrement() method in AtomicInteger gives the algorithm
for this:

public final int getAndIncrement() {
for (;;) {
int current = get();
int next = current + 1;
if (compareAndSet(current, next)) // calls native code
return current;
}
}

The native instruction to perform CAS guarantees that the operation
is atomic (all or nothing).

Public

5. Other Stuff
There are a few other small utilities that I think are quite useful:

CountDownLatch
• Allows one or more threads to wait at a certain point until
countdown is completed.
• Any thread that does the countdown will not wait at the point of
countdown. This is the difference between the CountDownLatch
and the CyclicBarrier.

CyclicBarrier
• Allows two or more threads to wait at a certain point until all
threads have arrived at the wait point. Then all threads will
resume execution at the same time.
• Useful when you want to sync up all threads to arrive at a certain
point in the code before proceeding further.

Public

5. Other Stuff (2)
Phaser (new in Java 7)
• Combines the functionality of both the CountDownLatch
and the CyclicBarrier.
• With the new Phaser, you will not need to use any of the
CountDownLatch and the CyclicBarrier anymore.

ThreadLocalRandom (new in Java 7)
• Better performance for Random when using multiple
threads.
• Implementation is only a ThreadLocal wrapper around
java.util.Random. Hence this is very useful in conceptual
reuseability.

Public

5. Other Stuff (3)
Semaphore
 Counter and waiter for availability of a particular
resource.

TimeUnit
 Time unit conversions utility

Public

Some Recommendations
 JDK collections are reasonably good quality
so you should use them and not re-invent
the wheel.
 Some open-source projects (Trove, Google
Collections) are worth looking at but do
expect only very slight performance gains.
 When in doubt, synchronize.
 But don’t over-synchronize unnecessarily!

5. Other Stuff - ThreadLocal
java.lang.ThreadLocal
• Not part of the concurrent package, but closely-
related to concurrency.
• Implementation is super-fast using a rehash
lookup hashtable (not linked list of entries).
• Saves programmers from needing to extend
java.lang.Thread to add additional attributes to
each thread.
• Entries for dead threads will be cleaned up by
GC automatically.

Public

5. Other Stuff - ThreadLocal Sample
• Don't add "synchronized" on method "initialValue")
• @Override is optional

private static final ThreadLocal<StringBuilder>
perThreadSB =
new ThreadLocal<StringBuilder>() {
@Override protected StringBuilder initialValue() {
return new StringBuilder();
}
};

Public

5. Other Stuff – Sample Code
Here is a simple example of how I often use barriers (using Phaser here). The
task is to run N threads concurrently, starting them at the same time, and
then stopping automatically after a few seconds.

SimpleBarrierExample.java:

import java.util.concurrent.Phaser;
import java.util.concurrent.ScheduledThreadPoolExecutor;
import java.util.concurrent.TimeUnit;

public class SimpleBarrierExample {
public static void main(String[] args) {
// Config.
int numThreads = 10;
int numSeconds = 3;

Public

// Init threads.
final Phaser p = new Phaser(numThreads + 1);
for (int i = 0; i < numThreads; i++) {
new Thread() {
public void run() {
p.arriveAndAwaitAdvance();
for (;;) doSomething();
}
}.start();
}

// Auto-exit after numSeconds seconds.
System.out.println("Wait " + numSeconds + " seconds");
new ScheduledThreadPoolExecutor(1).schedule(new Runnable() {
public void run() {
System.out.println("Done");
System.exit(0);
}
}, numSeconds, TimeUnit.SECONDS);

Public

// Start now.
p.arriveAndDeregister();
}

public static void doSomething() {
try {
System.out.println(Thread.currentThread().getId());
Thread.sleep(500L);
} catch (InterruptedException e) {
e.printStackTrace();
}
}

}

Public

End of Part II. Concurrent Utilities
We are done with a brief overview of what is
available out of the box from the
java.util.concurrent package.



Public

The End
Hope you had enjoyed this presentation, and now back to
coding work!!

Public

Abbreviations Used

CPU Central Processing Unit
JDK Java Development Kit
JMM Java Memory Model
JSR Java Specification Requests
JVM Java Virtual Machine

Public

References
All material presented here is publicly
available but this presentation is a
digested version of some selected online
materials. There is no reason why you
cannot go and dig up the same information
given enough time and motivation.

Hence this presentation is designed to save
you time.

Public

References - Official JSR-133
JSR-133: Java Memory Model and Thread
Specification
(Complete proposal document for JSR-133)
http://www.cs.umd.edu/~pugh/java/memoryModel
/jsr133.pdf

Java Language Specification – Section 17.4 Memory
Model
(The official documentation defining how the JMM works)
http://docs.oracle.com/javase/specs/jls/se7/html/jl
s-17.html#jls-17.4

Public

References - Low Level
The Unofficial JSR-133 Guide for Compiler Writers
(Human-readable guide to memory barriers required by JSR-133)
http://gee.cs.oswego.edu/dl/jmm/cookbook.html

CPU Caching
(General background info on how CPU caches work)
http://en.wikipedia.org/wiki/CPU_cache

Public

References - Unofficial JSR-133
FAQ to JSR 133 - The Java Memory Model
(Human-readable introduction to JSR-133)
http://www.cs.umd.edu/~pugh/java/memoryModel/jsr-
133-faq.html
Problems with the Old Java Memory Model
(Online article to summarize the old problems)
http://www.ibm.com/developerworks/library/j-
jtp02244/index.html
Fixes to the Old Java Memory Model
(Online article to summarize the new fixes)
http://www.ibm.com/developerworks/library/j-
jtp03304/

Public

References - Others
The "Double-Checked Locking is Broken"
Declaration
(Detailed description of the problems with double-checked locking)
http://www.cs.umd.edu/~pugh/java/memoryModel
/DoubleCheckedLocking.html

Using Unsafe in Java
(How to use C-pointers-style direct memory access in Java)
http://highlyscalable.wordpress.com/2012/02/02/di
rect-memory-access-in-java/

Public

Advanced Introduction to Java Multi-Threading - Full (chok)

More Related Content

What's hot (20)

Similar to Advanced Introduction to Java Multi-Threading - Full (chok) (20)

Recently uploaded (20)

Advanced Introduction to Java Multi-Threading - Full (chok)