我正在运行一个实验,以确定包装方法的性能开销。我读到JIT编译器和/或JVM优化了小方法,但我似乎一直遭受3-5%的性能罚款。
代码如下:
import java.util.* ;
public class WrappingTest1{
private WrappingTest1(){
// Empty.
}
private static void findPrimes(
final Long maxValue ,
final List< Long > foundPrimes
){
if(
maxValue > 2L
){
Boolean isPrime ;
foundPrimes.clear() ;
for(
Long i = 2L ;
i <= maxValue ;
i += 1L
){
isPrime = true ;
for(
Long j = 2L ;
j < i ;
j += 1L
){
if(
( i % j ) == 0
){
isPrime = false ;
}
}
if(
isPrime
){
foundPrimes.add(
i
) ;
}
}
}
}
private static void wrapper(
final Long input ,
final List< Long > output
){
findPrimes(
input ,
output
) ;
}
public static void main(
final String[] args
){
ArrayList< Long > primes ;
Long startTime ;
Long endTime ;
Double duration ;
Double meanDuration ;
Long primeRange ;
Long warmupIterations ;
Long benchmarkIterations ;
primes = new ArrayList<>() ;
meanDuration = 0.0 ;
primeRange = 100L ;
warmupIterations = 20000L ;
benchmarkIterations = 100000L ;
System.out.println(
"Experiment started."
) ;
// Unwrapped warmup.
for(
Long i = 0L ;
i < warmupIterations ;
i += 1L
){
findPrimes(
primeRange ,
primes
) ;
}
// Unwrapped benchmark.
startTime = System.nanoTime() ;
for(
Long i = 0L ;
i < benchmarkIterations ;
i += 1L
){
findPrimes(
primeRange ,
primes
) ;
}
endTime = System.nanoTime() ;
duration = ( endTime.doubleValue() - startTime.doubleValue() ) / 1E9 ;
System.out.println(
"Unwrapped runtime: " + duration + " seconds."
) ;
// Wrapped warmup.
for(
Long i = 0L ;
i < warmupIterations ;
i += 1L
){
wrapper(
primeRange ,
primes
) ;
}
// Wrapped benchmark.
startTime = System.nanoTime() ;
for(
Long i = 0L ;
i < benchmarkIterations ;
i += 1L
){
wrapper(
primeRange ,
primes
) ;
}
endTime = System.nanoTime() ;
duration = ( endTime.doubleValue() - startTime.doubleValue() ) / 1E9 ;
System.out.println(
"Wrapped runtime: " + duration + " seconds."
) ;
System.out.println(
"Experiment completed."
) ;
}
}
结果如下:
Experiment started.
Unwrapped runtime: 4.851473465 seconds.
Wrapped runtime: 5.078349508 seconds.
Experiment completed.
为什么会发生这种情况?如何使JVM内联方法或以其他方式对其进行优化以使包装器被忽略?
谢谢。
乐观,为迭代选择final Long maxValue
,但随后将100用作MaxValue。
如果您用ints更换long,则可能会得到10倍的速度。
第二和第三,急剧的改进是循环直到数学。
for (int j = 2; j <= Math.sqrt(i) && isPrime; ++j)
我知道,您并不是一个有效的Prime查找算法的背后,但这全都是针对包装器的微型计算标记,但是要学习的基本课程是,此类假设在大多数情况下不是一堆应用程序的瓶颈。
在此旁边,您应该在相反的方向上尝试测试,首先包裹,然后取消包装。为了简单地进行此类更改,您应该考虑循环和时间安排。
import java.util.* ;
public class WrappingTest1
{
PrimeFinder[] pfs = new PrimeFinder[2];
int primeRange = 1000;
private WrappingTest1 ()
{
// pfs[1] = new UnwrappedFinder ();
// pfs[0] = new WrappedFinder (pfs[1]);
pfs[0] = new UnwrappedFinder ();
pfs[1] = new WrappedFinder (pfs[0]);
}
void test ()
{
for (PrimeFinder pf: pfs)
runblock (pf);
}
void loopy (int iterations, PrimeFinder pf, ArrayList <Integer> primes)
{
for (int i = 0; i < iterations; ++i)
pf.findPrimes (primeRange, primes);
}
void runblock (PrimeFinder pf)
{
int warmupIterations = 20000;
int benchmarkIterations = 100000;
ArrayList <Integer> primes = new ArrayList<Integer> (50000) ;
// warmup.
loopy (warmupIterations, pf, primes);
// enchmark.
Long startTime = System.nanoTime();
loopy (benchmarkIterations, pf, primes);
Long endTime = System.nanoTime() ;
Double duration = (endTime.doubleValue () - startTime.doubleValue ()) / 1E9 ;
System.out.printf ("%s runtime: %4.2f seconds.n", pf.name(), duration);
// had to make sure, that we're really producing valid primes:
// and that they survive the code changes.
for (int p: primes) {
System.out.printf ("%d ", p);
}
System.out.println ("bye");
}
abstract class PrimeFinder {
abstract void findPrimes (final int maxValue, final List <Integer> foundPrimes);
abstract String name ();
}
class UnwrappedFinder extends PrimeFinder {
String name () {return "Unwrapped";}
void findPrimes (final int maxValue, final List <Integer> foundPrimes)
{
if (maxValue > 2)
{
foundPrimes.clear () ;
for (int i = 2; i <= maxValue; ++i)
{
Boolean isPrime = true;
for (int j = 2; j <= Math.sqrt(i) && isPrime; ++j)
if ((i % j) == 0)
isPrime = false;
if (isPrime)
foundPrimes.add (i);
}
}
}
}
class WrappedFinder extends PrimeFinder {
String name () {return " Wrapped";}
private PrimeFinder pf;
public WrappedFinder (PrimeFinder ppf)
{
pf = ppf;
}
void findPrimes (final int input, final List <Integer> output) {
pf.findPrimes (input, output);
}
}
public static void main (final String[] args)
{
System.out.println ("Experiment started.");
WrappingTest1 wt1 = new WrappingTest1 ();
wt1.test ();
System.out.println ("Experiment completed.") ;
}
}
使用Primerange = 100运行代码,但是1M迭代,我得到:
Unwrapped runtime: 2,34 seconds.
Unwrapped runtime: 2,53 seconds.
Unwrapped runtime: 2,50 seconds.
Unwrapped runtime: 2,50 seconds.
Unwrapped runtime: 2,49 seconds.
Unwrapped runtime: 2,52 seconds.
Unwrapped runtime: 2,59 seconds.
Unwrapped runtime: 2,60 seconds.
Unwrapped runtime: 2,58 seconds.
Unwrapped runtime: 2,52 seconds.
Wrapped runtime: 2,36 seconds.
Wrapped runtime: 2,36 seconds.
Wrapped runtime: 2,36 seconds.
Wrapped runtime: 2,36 seconds.
Wrapped runtime: 2,37 seconds.
Wrapped runtime: 2,37 seconds.
Wrapped runtime: 2,36 seconds.
Wrapped runtime: 2,41 seconds.
Wrapped runtime: 2,37 seconds.
Wrapped runtime: 2,37 seconds.
因此,令人惊讶的是,包装版本更快。嗯。更改PrimeFinder [] PFS中的顺序,它们更近,包裹着:2.46,未包装2.52。
看来,放缓的原因不是包装/未包装方法调用的优化(或缺乏(,而是缺乏对主方法本身的优化。用-XX进行分析:-printCompilation表明,根据默认设置(-XX:CompileTherShord = 10000(,包装/未包装的方法调用在10K热身迭代后进行了优化。但是,只有在大约70k的热身迭代之后,JVM分析报告了主要方法。因此,如果热身迭代小于70k,则未包装的运行时明显低于包装的运行时;但是,如果热身迭代为70k及以上,则两个跑步时间相似。当然,这仅适用于指定的基准测试 - 单个程序可能会产生不同的结果。