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

Question

I\'m trying to figure out how to correctly use Java\'s Executors. I realize submitting tasks to an ExecutorService has its own overhead. However, I\'m surpris

Answer 1

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...).