Performance Tuning Java Applications

Performance Tuning Java Applications Dave Scruggs Systems Engineer

Outline 1. 2.

General overview Suggested toolkit and procedures „ Demo 1—Temporary Objects

3.

Compilers and the JVM „ Demo 2—Memory Leaks

4.

Networking and I/O „ Demo 3—non-blocking I/O

5.

Threading „ Demo 4—CPU Bottlenecks

Know your target „

What kind of application are you writing? ƒ Embedded, Workstation, Workgroup, Intranet, or Enterprise?

„

ƒ Personalized or Generalized? What is your user audience?

„

ƒ Technically Sophisticated? ƒ Impatient? What is your data like? ƒ Detailed, hard to generate reports or simple data? ƒ Persistent or volatile? ƒ Order-entry or decision support?

„

How are you interacting? ƒ Synchronous or asynchronous?

Why doesn’t my application perform? „

„

Political Constraints (Boss, Client requires it “done a certain way) Infrastructure—both the infrastructure you have and what you don’t. ƒ Memory, CPU, OS, Disk Space ƒ Directory Servers, Databases, Legacy Applications ƒ Network ƒ Application Server

„ „

Time—you have limited resources Money

„

Third-party technologies JVM and/or JDK

„

Compiler

„

How It Doesn’t Perform „

„ „ „

Scalability: “It worked great for the first 50 users…” ƒ CPU Utilization ƒ Memory Utilization ƒ I/O Utilization Reliability: “Where did it go?” Latency: “It works great, once it gets started…” Throughput: “How long, on average, does it take to do ‘x’?” ƒ Algorithm Design ƒ Architecture

What to Do… „

„

Political Constraints: ƒ Work within constraints ƒ get a friend who knows how to work these situations Infrastructure: ƒ Test software on the systems you’ll be using ƒ Identify weak portions of technology and avoid them ƒ Tune your code

„ „ „

„

Time—get good tools, choose methodology well Money—see “Political Constraints” Third-party technologies ƒ Tune your code JVM, JDK, and Compiler—choose faster option, if supported and available

A Tuner’s Toolkit „ „ „ „

Packet Sniffer Profiler (Code coverage, Memory profiler, Thread profiler) Unit tests Load tester ƒ JMeter ƒ java.awt.Robot ƒ Expect (if you can use Tcl)

„ „

ƒ Mercury, Silk, other commercial packages Network, Architecture, and Class Diagrams Other environments (OS, Foreign languages, AppServer, VMWare, Browsers…)

Performance Testing Workflow start

analysis

design

develop unit test

no

O.K. yes

perf test profile

no

perf o.k.?

yes

stop

Temporary Objects Profiling Demo

What Is a JVM? „

„

„

Can be thought of as a “processor on top of a processor.” ƒ Has a program counter, stacks, method memory space, and heap space. ƒ Direct support for the following types: int, long, float, double, “references”, and return addresses (hidden, no pointers). ƒ Partial support for boolean, byte, char, and short—partially treated as integers. ƒ Return values are kept in “frames”, which are allocated and deallocated as methods run and complete. Distinct from the class libraries supplied with the JDK or java runtime. Native methods throw a “twist” into how the JVM “hardware” values are interpreted--native method “stacks” may be handled differently by different VM’s (or not supported at all).

JVM Diagram

Execution Engine (“Clock”, ALU, ...)

pc

Thread-local native stack

stack

…

…

…

…

…

…

ir

heap

bne

…

…

…

dup

…

…

…

…

…

…

…

…

…

JVM Characteristics to Watch Out for „

„

„

Many JVM’s (including sun’s) allow the heap to grow. Although garbage collection frees space in the heap, the heap itself doesn’t shrink. Consequently, creating a lot of objects can steal memory away from other processes that can’t be reclaimed without killing a java process. There are two “sizes” of objects that go on the stack—int width and long width. Variables of long and double width take up two entries on the stack, and take twice as long to hop on or off. The short, boolean, byte, references, and char variables take up an int entry—the rest of the space (if any) is wasted. Frames hold local variables, incoming operands, and return values. Again, doubles and longs take up twice as much room.

A Very Simple Class package hello; /** * This is a very simple class, just to show what it takes takes * to run a simple method. */ public class SimpleHello { // private attributes private String message; private String username; // Constructors public SimpleHello() SimpleHello() { message = "Hello"; username = "Dude"; } public void setMessage(String setMessage(String message) { this.message = message; } public String getMessage() getMessage() { return message; }

public void setUsername(String setUsername(String username) { this.username = username; } public String getUsername() getUsername() { return username; } public void sayHello(){ sayHello(){ System.out.println (message + " " + username + "!"); System.out.println(message } public long addTwoLong(long addTwoLong(long a, long b){ return a + b; } public int addTwoInt( addTwoInt(int a, int b){ return a + b; } public static void main(String[] args) args) { SimpleHello simpleHello1 = new SimpleHello(); SimpleHello(); simpleHello1.sayHello (); simpleHello1.sayHello(); } }

How ““SimpleHello” SimpleHello” Appears to the VM SimpleHello Constructor 0 aload_0 1 invokespecial #1 <Method java.lang.Object()> 4 aload_0 5 ldc #2 <String "Hello"> 7 putfield #3 10 aload_0 11 ldc #4 <String "Dude"> 13 putfield #5 16 return

SimpleHello Constructor public SimpleHello() { message = "Hello"; username = "Dude"; }

addTwoLong

addTwoLong

0 lload_1 1 lload_3 2 ladd 3 lreturn Local variables for method long addTwoLong(long, long) long b pc=0, length=4, slot=3 long a pc=0, length=4, slot=1 hello.SimpleHello this pc=0, length=4, slot=0

public long addTwoLong(long addTwoLong(long a, long b){ return a + b; }

How to “Say Hello” sayHello 0 getstatic #6 3 new #7 6 dup 7 aload_0 8 getfield #3 11 invokestatic #8 <Method java.lang.String valueOf(java.lang.Object)> 14 invokestatic #8 <Method java.lang.String valueOf(java.lang.Object)> 17 invokespecial #9 <Method java.lang.StringBuffer(java.lang.String)> 20 ldc #10 <String " "> 22 invokevirtual #11 <Method java.lang.StringBuffer append(java.lang.String)> 25 aload_0 26 getfield #5 29 invokevirtual #11 <Method java.lang.StringBuffer append(java.lang.String)> 32 ldc #12 <String "!"> 34 invokevirtual #11 <Method java.lang.StringBuffer append(java.lang.String)> 37 invokestatic #8 <Method java.lang.String valueOf(java.lang.Object)> 40 invokestatic #8 <Method java.lang.String valueOf(java.lang.Object)> 43 invokevirtual #13 <Method void println(java.lang.String)> 46 return

sayHello public void sayHello(){ sayHello(){ System.out.println (message + " " + username + "!"); System.out.println(message }

Compilers „ „

„

„

Javac BCJ: JBuilder’s compiler—optimizations, including obfuscation, targeting specific releases of the JVM, and controlling the amount of debug information provided within the classes. Jikes: Provides some optimization, and a very useful “lint” feature University of Indiana’s “JAVAR”—finds parallelism in code

Memory Leak Profiling Demo

Networking and I/O „ „

„

„

„ „

JDK 1.4 has improved I/O classes (nonblocking I/O, channels) Nonblocking I/O is available for earlier JDK’s at http://www.cs.berkeley.edu/~mdw/proj/java-nbio/ Watch FileInputStream, FileOutputStream, and RandomAccessFile. (FileReader and FileWriter use the streams) Be careful passing collections of objects through RMI/IIOP or CORBA—these objects are passed as “any’s”, which are heavy. Minimize objects transferred—transfer only what you have to. Speed generally follows this order: 1. Register Access and Cache (local primitive variables) 2. RAM (heap variables, arrays, and all classes) 3. Disk (files) 4. Network (remote calls)

How Do I Cut Down Network Traffic? „ „ „

„

„

„ „

Keep a local (cached) copy. Value Objects work well, here. Choose your technology wisely (CORBA and RMI beat SOAP). Message Queues can help when the data transfer doesn’t require synchronous connections—they won’t block. Identify tight communication “communities”, and colocate that code. Use value objects for getter’s and setter’s—and reduce the network traffic from a number of small calls to one big one. Validate input data as close to the client as possible. Use constants where possible (for instance “1” is easier that “Could not get a database connection, stack trace…”).

What About Databases? „

Use prepared statements, when possible. Don’t use “SELECT *”.

„

Use the lightest transaction possible

„

1. TRANSACTION_NONE 2. TRANSACTION_READ_UNCOMMITTED 3. TRANSACTION_READ_COMMITTED 4. TRANSACTION_REPEATABLE_READ 5. TRANSACTION_SERIALIZABLE

„

ODBMS systems work better for pure OO data. Use Connection Pools.

„

Batch queries, if possible.

„

Non-blocking I/O Demo

Threads „

„

„

Even though two threads can theoretically synchronize at the same time, the mutex locking process itself is single-threaded. Prefer Maps to Hashtables, and ArrayLists to Arrays, since the former allow a choice of synchronization. Parallelizing work between multiple threads can help ƒ If you can, keep as much code as you can within thread-local storage, and “synchronize at the end” ƒ Thread pools can save work for heavily loaded applications.

„ „

Produce working, thread-safe code first—then optimize. Synchronization on its own is not enough to provide proper code—threads have to act in the proper order.

Collection Synchronization Vector synchronizes the entire object on these methods: add() addAll() addAll() clone() containsAll() containsAll() elements().nextElement () elements().nextElement() equals() get() hashCode() hashCode() remove() (calls "removeElement ") "removeElement") removeAll() removeAll() removeAllElements() removeAllElements() set() toArray() toArray() toString() toString()

Hashtable synchronizes the entire object on these methods: clear() clone() contains(Object obj) obj) containsKey(Object containsKey(Object obj) obj) elements() equals(Object obj) obj) get(Object obj) obj) hashCode() hashCode() keys() put(Object obj, obj, Object obj1) putAll(Map putAll(Map map) readObject( readObject(ObjectInputStream objectinputstream) objectinputstream) remove() [synchronized(Hashtable .this)] [synchronized(Hashtable.this)] remove(Object obj) obj) toString() toString() writeObject( writeObject(ObjectOutputStream objectoutputstream) objectoutputstream)

If you use a vector or hashtable for configuration data, you’re blocking other parts of your application.

Execution Order „ „

Actions performed by one thread are ordered. For any single variable on the heap, access is implicitly synchronized . ƒ If two threads write to the same variable, one always writes before the other. ƒ Which thread accesses a heap variable first is undetermined

„

ƒ WAW, WAR, and RAW hazards are still possible. Within a monitored (locked) section, access is synchronized.

„

Actions only happen once.

CPU Profiling Demo

Bibliography and References „

„

„

„

„

„

„

Shirazi, Jack, Java Performance Tuning, O'Reilly and Associates, Sebastapol, CA, September 2000 Wilson, Steve and Jeff Kesselman, Java Platform Performance: Strategies and Tactics (First Edition), Addison Wesley, Reading, MA, http://java.sun.com/docs/books/performance/ Lea, Doug, Concurrent Programming in Java, Addison Wesley, Reading, MA, 1997 Lindholm, Tim and Frank Yellin, The Java Virtual Machine Specification Second Edition, Addison Wesley, Reading, MA,1999 Venners, Bill, Inside the Java Virtual Machine, McGraw-Hill Professional Publishing, December 1999 Alur, Deepak and John Crupi and Dan Malks, Core J2EE Patterns: Best Practices and Design Strategies, Prentice Hall PTR, Palo Alto, CA,2001 Hennessy, John and David Patterson, Computer Architecture: A Quantitative Approach, Morgan Kaufmann Publishers, Palo Alto, CA, 1990