ExecutorService's surprising performance break-even point --- rules of thumb?

Answer 1

This is not a fair test for the thread pool for following reasons,

1. You are not taking advantage of the pooling at all because you only have 1 thread.
2. The job is too simple that the pooling overhead can't be justified. A multiplication on a CPU with FPP only takes a few cycles.

Considering following extra steps the thread pool has to do besides object creation and the running the job,

1. Put the job in the queue
2. Remove the job from queue
3. Get the thread from the pool and execute the job
4. Return the thread to the pool

When you have a real job and multiple threads, the benefit of the thread pool will be apparent.

Answer 2

Firstly there's a few issues with the microbenchmark. You do a warm up, which is good. However, it is better to run the test multiple times, which should give a feel as to whether it has really warmed up and the variance of the results. It also tends to be better to do the test of each algorithm in separate runs, otherwise you might cause deoptimisation when an algorithm changes.

The task is very small, although I'm not entirely sure how small. So number of times faster is pretty meaningless. In multithreaded situations, it will touch the same volatile locations so threads could cause really bad performance (use a Random instance per thread). Also a run of 47 milliseconds is a bit short.

Certainly going to another thread for a tiny operation is not going to be fast. Split tasks up into bigger sizes if possible. JDK7 looks as if it will have a fork-join framework, which attempts to support fine tasks from divide and conquer algorithms by preferring to execute tasks on the same thread in order, with larger tasks pulled out by idle threads.

Answer 3

1. Using executors is about utilizing CPUs and / or CPU cores, so if you create a thread pool that utilizes the amount of CPUs at best, you have to have as many threads as CPUs / cores.
2. You are right, creating new objects costs too much. So one way to reduce the expenses is to use batches. If you know the kind and amount of computations to do, you create batches. So think about thousand(s) computations done in one executed task. You create batches for each thread. As soon as the computation is done (java.util.concurrent.Future), you create the next batch. Even the creation of new batches can be done in parralel (4 CPUs -> 3 threads for computation, 1 thread for batch provisioning). In the end, you may end up with more throughput, but with higher memory demands (batches, provisioning).

Edit: I changed your example and I let it run on my little dual-core x200 laptop.

provisioned 2 batches to be executed
simpleCompuation:14
computationWithObjCreation:17
computationWithObjCreationAndExecutors:9

As you see in the source code, I took the batch provisioning and executor lifecycle out of the measurement, too. That's more fair compared to the other two methods.

See the results by yourself...

import java.util.List;
import java.util.Vector;
import java.util.concurrent.ExecutorService;
import java.util.concurrent.Executors;
import java.util.concurrent.TimeUnit;

public class ExecServicePerformance {

    private static int count = 100000;

    public static void main( String[] args ) throws InterruptedException {

        final int cpus = Runtime.getRuntime().availableProcessors();

        final ExecutorService es = Executors.newFixedThreadPool( cpus );

        final Vector< Batch > batches = new Vector< Batch >( cpus );

        final int batchComputations = count / cpus;

        for ( int i = 0; i < cpus; i++ ) {
            batches.add( new Batch( batchComputations ) );
        }

        System.out.println( "provisioned " + cpus + " batches to be executed" );

        // warmup
        simpleCompuation();
        computationWithObjCreation();
        computationWithObjCreationAndExecutors( es, batches );

        long start = System.currentTimeMillis();
        simpleCompuation();
        long stop = System.currentTimeMillis();
        System.out.println( "simpleCompuation:" + ( stop - start ) );

        start = System.currentTimeMillis();
        computationWithObjCreation();
        stop = System.currentTimeMillis();
        System.out.println( "computationWithObjCreation:" + ( stop - start ) );

        // Executor

        start = System.currentTimeMillis();
        computationWithObjCreationAndExecutors( es, batches );    
        es.shutdown();
        es.awaitTermination( 10, TimeUnit.SECONDS );
        // Note: Executor#shutdown() and Executor#awaitTermination() requires
        // some extra time. But the result should still be clear.
        stop = System.currentTimeMillis();
        System.out.println( "computationWithObjCreationAndExecutors:"
                + ( stop - start ) );
    }

    private static void computationWithObjCreation() {

        for ( int i = 0; i < count; i++ ) {
            new Runnable() {

                @Override
                public void run() {

                    double x = Math.random() * Math.random();
                }

            }.run();
        }

    }

    private static void simpleCompuation() {

        for ( int i = 0; i < count; i++ ) {
            double x = Math.random() * Math.random();
        }

    }

    private static void computationWithObjCreationAndExecutors(
            ExecutorService es, List< Batch > batches )
            throws InterruptedException {

        for ( Batch batch : batches ) {
            es.submit( batch );
        }

    }

    private static class Batch implements Runnable {

        private final int computations;

        public Batch( final int computations ) {

            this.computations = computations;
        }

        @Override
        public void run() {

            int countdown = computations;
            while ( countdown-- > -1 ) {
                double x = Math.random() * Math.random();
            }
        }
    }
}

Answer 4

I don't think this is at all realistic since you're creating a new executor service every time you make the method call. Unless you have very strange requirements that seems unrealistic - typically you'd create the service when your app starts up, and then submit jobs to it.

If you try the benchmarking again but initialise the service as a field, once, outside the timing loop; then you'll see the actual overhead of submitting Runnables to the service vs. running them yourself.

But I don't think you've grasped the point fully - Executors aren't meant to be there for efficiency, they're there to make co-ordinating and handing off work to a thread pool simpler. They will always be less efficient than just invoking Runnable.run() yourself (since at the end of the day the executor service still needs to do this, after doing some extra housekeeping beforehand). It's when you are using them from multiple threads needing asynchronous processing, that they really shine.

Also consider that you're looking at the relative time difference of a basically fixed cost (Executor overhead is the same whether your tasks take 1ms or 1hr to run) compared to a very small variable amount (your trivial runnable). If the executor service takes 5ms extra to run a 1ms task, that's not a very favourable figure. If it takes 5ms extra to run a 5 second task (e.g. a non-trivial SQL query), that's completely negligible and entirely worth it.

So to some extent it depends on your situation - if you have an extremely time-critical section, running lots of small tasks, that don't need to be executed in parallel or asynchronously then you'll get nothing from an Executor. If you're processing heavier tasks in parallel and want to respond asynchronously (e.g. a webapp) then Executors are great.

Whether they are the best choice for you depends on your situation, but really you need to try the tests with realistic representative data. I don't think it would be appropriate to draw any conclusions from the tests you've done unless your tasks really are that trivial (and you don't want to reuse the executor instance...).