Re: [boost] [threadpool] version 22 with default pool

Subject: Re: [boost] [threadpool] version 22 with default pool
From: Edouard A. (edouard_at_[hidden])
Date: 2009-03-08 15:38:22

> Am Sunday 08 March 2009 19:48:11 schrieb Edouard A.:
> > I have implemented a parallel_sort that is a bit slower than tbb with
> > threadpool (see below for the code).
> threadpool uses a signle-lock global queue - if a thread enters the
> queue for
> enqueing/dequeing a task it aquires the lock so no other thread can
> enter the
> queue. with a lock-free implementation it should work faster.

This is quite possible. I'm looking forward to test it with a lock free
queue. I've also seen a lots of .lock() .unlock() in the code... Sometimes
holding the lock is faster than releasing it and acquiring it again.

> pool::shutdown() is not correct - use boost::wait_for_all() or
> boost::wait_for_any() from the future lib from Anthony WIlliams (I've
> added
> the lib to threadpool archive).
> 
> task< int > tsk1 = pool.submit(...);
> task< string > tsk2 = pool.submit(...);
> task< int > tsk3 = pool.submit(...);
> 
> wait_for_all( tsk1.result(), tsk2.result(), tsk3.result());
> // here all tasks are finsihed

I'm not sure it's a good idea to force the user to collect the tasks and
wait for them. I thought threadpool could abstract that out. I mean, it's
nice to have the alternative to wait for specific tasks, but generally you
just throw in work into the pool and want that work to be finished...
Perhaps a task_group object could solve this problem? You could also "link"
tasks. Ie, you need depending tasks to be finished when you wait on one.

Anyway, with wait_for_all, it's clearly slower than with shutdown(). Not
really surprising as you go through more abstraction layers and you have to
store the tasks' results in a structure. IMHO, the fastest way would
probably to have a condition_variable bound to the number of running tasks
in the pool itself.

With wait_for_all:

                    std::fill 0..1000000 ok : 1 elapsed: 0.01
       tbb::parallel_for fill 0..1000000 ok : 1 elapsed: 0.01
            std::sort reverse 0..1000000 ok : 1 elapsed: 0.111
   tbb::parallel_sort reverse 0..1000000 ok : 1 elapsed: 0.038
      boost::tp::sort reverse 0..1000000 ok : 1 elapsed: 0.055

Regards.

-- 
EA