Atomic Counters in Cilk
Here’s the translated code and explanation in Markdown format suitable for Hugo:
The primary mechanism for managing state in many languages is communication over channels or message passing. However, there are a few other options for managing state. Here we’ll look at using atomic operations for atomic counters accessed by multiple threads.
import java.util.concurrent.atomic.AtomicLong;
import java.util.concurrent.CountDownLatch;
public class AtomicCounters {
public static void main(String[] args) throws InterruptedException {
// We'll use an AtomicLong to represent our (always-positive) counter.
AtomicLong ops = new AtomicLong();
// A CountDownLatch will help us wait for all threads to finish their work.
CountDownLatch latch = new CountDownLatch(50);
// We'll start 50 threads that each increment the counter exactly 1000 times.
for (int i = 0; i < 50; i++) {
new Thread(() -> {
for (int c = 0; c < 1000; c++) {
// To atomically increment the counter we use incrementAndGet().
ops.incrementAndGet();
}
latch.countDown();
}).start();
}
// Wait until all the threads are done.
latch.await();
// Here no threads are writing to 'ops', but using get() it's safe to
// atomically read a value even while other threads are (atomically) updating it.
System.out.println("ops: " + ops.get());
}
}
We expect to get exactly 50,000 operations. Had we used a non-atomic integer and incremented it with ops++
, we’d likely get a different number, changing between runs, because the threads would interfere with each other. Moreover, we’d get data race failures when running with proper concurrency analysis tools.
$ javac AtomicCounters.java
$ java AtomicCounters
ops: 50000
Next we’ll look at locks, another tool for managing state.