I am aggregating multiple values for keys in a multi-threaded environment. The keys are not known in advance. I thought I would do something like this:
class
The approach with putIfAbsent has the fastest execution time, it is from 2 to 50 times faster than the "lambda" approach in evironments with high contention. The Lambda isn't the reason behind this "powerloss", the issue is the compulsory synchronisation inside of computeIfAbsent prior to the Java-9 optimisations.
the benchmark:
import java.util.Random;
import java.util.concurrent.ConcurrentHashMap;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;
import java.util.concurrent.atomic.AtomicInteger;
import java.util.concurrent.atomic.AtomicLong;
public class ConcurrentHashMapTest {
private final static int numberOfRuns = 1000000;
private final static int numberOfThreads = Runtime.getRuntime().availableProcessors();
private final static int keysSize = 10;
private final static String[] strings = new String[keysSize];
static {
for (int n = 0; n < keysSize; n++) {
strings[n] = "" + (char) ('A' + n);
}
}
public static void main(String[] args) throws InterruptedException {
for (int n = 0; n < 20; n++) {
testPutIfAbsent();
testComputeIfAbsentLamda();
}
}
private static void testPutIfAbsent() throws InterruptedException {
final AtomicLong totalTime = new AtomicLong();
final ConcurrentHashMap map = new ConcurrentHashMap();
final Random random = new Random();
ExecutorService executorService = Executors.newFixedThreadPool(numberOfThreads);
for (int i = 0; i < numberOfThreads; i++) {
executorService.execute(new Runnable() {
@Override
public void run() {
long start, end;
for (int n = 0; n < numberOfRuns; n++) {
String s = strings[random.nextInt(strings.length)];
start = System.nanoTime();
AtomicInteger count = map.get(s);
if (count == null) {
count = new AtomicInteger(0);
AtomicInteger prevCount = map.putIfAbsent(s, count);
if (prevCount != null) {
count = prevCount;
}
}
count.incrementAndGet();
end = System.nanoTime();
totalTime.addAndGet(end - start);
}
}
});
}
executorService.shutdown();
executorService.awaitTermination(Long.MAX_VALUE, TimeUnit.DAYS);
System.out.println("Test " + Thread.currentThread().getStackTrace()[1].getMethodName()
+ " average time per run: " + (double) totalTime.get() / numberOfThreads / numberOfRuns + " ns");
}
private static void testComputeIfAbsentLamda() throws InterruptedException {
final AtomicLong totalTime = new AtomicLong();
final ConcurrentHashMap map = new ConcurrentHashMap();
final Random random = new Random();
ExecutorService executorService = Executors.newFixedThreadPool(numberOfThreads);
for (int i = 0; i < numberOfThreads; i++) {
executorService.execute(new Runnable() {
@Override
public void run() {
long start, end;
for (int n = 0; n < numberOfRuns; n++) {
String s = strings[random.nextInt(strings.length)];
start = System.nanoTime();
AtomicInteger count = map.computeIfAbsent(s, (k) -> new AtomicInteger(0));
count.incrementAndGet();
end = System.nanoTime();
totalTime.addAndGet(end - start);
}
}
});
}
executorService.shutdown();
executorService.awaitTermination(Long.MAX_VALUE, TimeUnit.DAYS);
System.out.println("Test " + Thread.currentThread().getStackTrace()[1].getMethodName()
+ " average time per run: " + (double) totalTime.get() / numberOfThreads / numberOfRuns + " ns");
}
}
The results:
Test testPutIfAbsent average time per run: 115.756501 ns
Test testComputeIfAbsentLamda average time per run: 276.9667055 ns
Test testPutIfAbsent average time per run: 134.2332435 ns
Test testComputeIfAbsentLamda average time per run: 223.222063625 ns
Test testPutIfAbsent average time per run: 119.968893625 ns
Test testComputeIfAbsentLamda average time per run: 216.707419875 ns
Test testPutIfAbsent average time per run: 116.173902375 ns
Test testComputeIfAbsentLamda average time per run: 215.632467375 ns
Test testPutIfAbsent average time per run: 112.21422775 ns
Test testComputeIfAbsentLamda average time per run: 210.29563725 ns
Test testPutIfAbsent average time per run: 120.50643475 ns
Test testComputeIfAbsentLamda average time per run: 200.79536475 ns