Subject: Re: [Boost-users] [asio] io_service threadpool performance
From: Oliver Seiler (oseiler_at_[hidden])
Date: 2009-03-17 16:39:50
On Fri, Mar 13, 2009 at 2:32 PM, Rutger ter Borg <rutger_at_[hidden]> wrote:
>
> Dear all,
>
> I have been testing asio's io_service in a threadpool setup for job
> dispatching. However, it seems as if adding threads doesn't improve
> performance; perhaps even the opposite with 1 thread having the best
> performance.
Someone more familiar with the implementation could comment, but just poking
around through the implementation it appears there is one queue of handlers
that will get shared by all threads; right there I think there'll be a
lot of lock contention
between threads on the single queue.
I tried translating this example to Intel's TBB library, and start to
see concurrency
effects as I move up beyond 8 threads on my quad-core box (using parallel_for
with a blocked_range that results in a single call to f per task).
Increasing the
amount of work done on a given task (by increasing the size of the blocked_range
to parallel_for) speeds the run-time greatly, presumably because of the reduced
number of tasks and reduced context switching.
I'm guessing that the asio io_service isn't really geared towards
effective use of
multi-core CPUs where you're trying to schedule a large number of small
computational tasks; I'll go out on a limb and say that this *wasn't* the intent
of the library (as the name somewhat implies).
Not sure if that was helpful, but it let me play around with TBB,
which seems very
nice.
Cheers
Oliver