JavaMicroBenchmarkpptm

Java Benchmarking
Srinivasan Raghavan
Senior Member of Technical Staff
Java Platform Group
Copyright © 2015, Oracle and/or its affiliates. All rights reserved. |

Program Agenda
Why do we need to benchmark ?
Do we benchmark correctly ?
Know the optimization ..
How open jmh works ..
Advanced topics ..
1
2
3
4
4
5

Why do we need to benchmark?
5

Trends in hardware tech
• Single processor performance improvement have slowing steadily from 2003
• Clock speed have not increased last decade due to power consumption factors
• Also there is a concern in microprocessor production tech which is hitting 7 nm process
barrier
• DRAM chip capacity has increased by about 25% to 40% per year recently. And there is
also tremendous increase in bandwidth. But latency is still a concern
• Its increasing difficulty of efficiently manufacturing even smaller DRAM cells
• Bandwidth has outpaced latency across these technologies and will likely continue to do
so
6

Why do you need to measure?
• Software Engineering is more like the Interstellar movie . Can be realistic, driven by
math. Its still a movie not a reality
• Performance Engineering is firmly placed in reality where one has to deal complex
hardware interaction, compiler and hardware optimization and multithreading
• The aim of Performance Engineering is to gather the performance model of the
underlying system
• It can give a picture where optimization is required and where there is too much over
engineering
• Having a microbench mark data is a far better that writing code in the blind
7

Do we benchmark correctly ?
8

A wrong approach to benchmark
9
interface Incrementer {
void increment();
}
class LockIncrementer implements Incrementer {
private long counter = 0;
private Lock lock = new ReentrantLock();
public void increment() {
lock.lock();
try {
++counter;
} finally {
lock.unlock();
}}}
class SyncIncrementer implements Incrementer {
private long counter = 0;
public synchronized void increment() {
++counter;
}

A wrong approach to benchmark
10
long test(Incrementer incr) {
long start = System.nanoTime();
for (long i = 0; i < 10000000L; i++)
incr.increment();
return System.nanoTime() - start;
}
public static void main(String[] args) {
long synchTime = test(new SyncIncrementer());
long lockTime = test(new LockIncrementer());
System.out.printf("synchronized: %1$10dn", synchTime);
System.out.printf("Lock: %1$10dn", lockTime);
System.out.printf("Lock/synchronized = %1$.3f", (double) lockTime /
(double) synchTime);
}

What are things missed here ?
• No consideration for compiler and hardware optimization and number of cores etc
• No consideration for the JVM optimization
• No consideration for the number of threads acting here.
• No consideration for variation of number of inputs
• Conclusion solely based on numbers and stack overflow conversation
11

Open JMH
• JMH is a Java harness for building, running, and analyzing nano/micro/milli/macro
benchmarks written in Java and other languages targeting the JVM.
• Part of the code-tools project of openjdk
• Used extensively within open jdk to test the internals
• Keeps pace with the changes in the jvm
• Brings scientific approach to benchmarking
12

A quick look at JMH working
• A maven based project . Bundles the benchmark code with the working jar
• A quick common annotation list
13
Annotation Function
@Benchmark Lines up the method for benchmarking
@BenchmarkMode Defines mode of benchmark line
averagetime or throughput
@Warmup Defines the warm-up cycles
@Measurement Defines the measurement iteration
@Fork Number of vm

Know the optimization
14

VM and Hardware optimization
• Dead code elimination – Elimination of code which is not used
• Inlining – Analyzing the outcome of the code and optimizing it
• Loop unrolling – increase a program's speed by reducing (or eliminating) instructions
that control the loop
• Warmup – VM starts by interpreting the code and after seeing the hot methods it starts
aggressive inlining
15

Copyright © 2015, Oracle and/or its affiliates. All rights reserved. | 16
@Benchmark
public void testMethod() {
int sum = 0;
for (int i = 0; i < 50; i++) {
sum += i;
}
}
@Benchmark
public int testMethod_1() {
int sum = 0;
for (int i = 0; i < 50; i++) {
sum += i;
}
return sum;
}
Benchmark Mode Cnt Score Error Units
Benchmark_Inlining.testMethod avgt 5 0.411 ▒ 0.199 ns/op
Benchmark_Inlining.testMethod_1 avgt 5 3.396 ▒ 0.138 ns/op
Benchmark_Inlining.testMethod_2 avgt 5 5.123 ▒ 0.993 ns/op
@CompilerControl(Mode.DONT_INLINE)
@Benchmark
public int testMethod_2() {
int sum = 0;
for (int i = 0; i < 50; i++) {
sum += i;
}
return sum;
}
DCE and Inlining

private double[] A = new double[2048];
@Benchmark
public double test1() {
double sum = 0.0;
for (int i = 0; i < A.length; i++) {
sum += A[i];
}
return sum;
}
@Benchmark
public double testManualUnroll() {
double sum = 0.0;
for (int i = 0; i < A.length; i += 4) {
sum += A[i] + A[i + 1] + A[i + 2] + A[i + 3];
}
return sum;
}
Benchmark_LoopUnroll.test1 avgt 5 1946.006 ▒ 73.579 ns/op
Benchmark_LoopUnroll.testManualUnroll avgt 5 823.572 ▒ 183.984 ns/op
Loop Unrolling

JMHSample_11_Loops.measure avgt 5 5.472 ▒ 2.783 ns/op
JMHSample_11_Loops.measure1 avgt 5 4.813 ▒ 0.352 ns/op
Unrolling and Inline in Steroids
int x = 1;
int y = 2;
@Benchmark
public int measure() {
return (x + y);
}
private int reps(int reps) {
int s = 0;
for (int i = 0; i < reps; i++) {
s += (x + y);
}
return s;
}
@Benchmark
@OperationsPerInvocation(1)
public int measure1() {
return reps(1);
}
@Benchmark
return reps(10000);
}
@Benchmark
return reps(100000);
}

Advanced Topics
19

BenchmarkAtomicInteger.baseline avgt 5 4.109 ▒ 0.345 ns/op
BenchmarkAtomicInteger.incrPlain avgt 5 3.320 ▒ 0.415 ns/op
BenchmarkAtomicInteger.incrAtomic avgt 5 9.205 ▒ 1.031 ns/op
What can be the cost of Atomic write
private int plainV;
private AtomicInteger atomicInteger = new
AtomicInteger(0);
@Benchmark
public int baseline() {
return 42;
}
@Benchmark
public int incrPlain() {
return plainV++;
}
@Benchmark
public int incrAtomic() {
return atomicInteger.incrementAndGet();
}

Benchmark (tokens) Mode Cnt Score Error Units
BenchmarkAmortizedAtomicInteger.baseline 40 avgt 5 84.652 ▒ 3.229 ns/op
BenchmarkAmortizedAtomicInteger.incrPlain 40 avgt 5 85.101 ▒ 2.916 ns/op
BenchmarkAmortizedAtomicInteger.incrAtomic 40 avgt 5 87.118 ▒ 3.891 ns/op
Lets Amortize the cost
@Param({ "40" })
private int tokens;
private int plainV;
private AtomicInteger atomicInteger = new
AtomicInteger(0);
@Benchmark
public int baseline() {
Blackhole.consumeCPU(tokens);
return 42;
}
@Benchmark
public int incrPlain() {
return plainV++;
}
@Benchmark
public int incrAtomic() {
return atomicInteger.incrementAndGet();
}

Conclusion
• Benchmark for understand the performance model of the system not obtain number for
fighting in stack over flow.
• It can give an insight to where performance tweaking is need and where its not required
where underlying systems can do the optimization for you
• Superficial conclusion without accurate measurement on performance can lead over
engineering.
22

References
• http://hg.openjdk.java.net/code-tools/jmh/
• http://openjdk.java.net/projects/code-tools/jmh/
• http://shipilev.net/
• Computer Architecture A Quantitative Approach (5th edition) John L. Hennessy
Stanford University David A. Patterson University of California, Berkeley
• http://openjdk.java.net/projects/jdk8/
23

Questions ?
24

JavaMicroBenchmarkpptm

More Related Content

What's hot (20)

Viewers also liked (20)

Similar to JavaMicroBenchmarkpptm (20)

JavaMicroBenchmarkpptm

Editor's Notes