Re: [boost] [threadpool] mixture of recursive invokation and fibers?

Subject: Re: [boost] [threadpool] mixture of recursive invokation and fibers?
From: k-oli_at_[hidden]
Date: 2008-11-05 14:01:32

Am Mittwoch, 5. November 2008 16:11:38 schrieb vicente.botet:
> In my opinion the scoped constructions do not work well to configure the
> threadpool, as we have no multiple threads that can configure the same
> object.

 noticed that in one implementation

> I don't know if both can be configured at the same time.
>
> At least we have two variables to configure:
>     task submision : where a task will be enqueued (pool queue or worker
> thread queue)
>     blocking behaviour: what the scheduler does when we block e.g. on get
> (blocks, schedule recursilvely, use fibers, ...)
>
> Concerning the working thread in addition we could consider whether other
> working thread schedulers can steal tasks for this working thread queue.
>
> I'm not sure if we need all these variations, but it would be better to
> separate the concerns and see how this work. We need to see how all
> variables interacts and which ones are compatible and/or desirables. For
> example if the task is submited on the internal queue and the scheculer
> blocks on blocking functions these task will never run.
>
> Your first thread_pool uses the pool thread, don't have internal queue and
> don't steal from others working threads, and of course do not allow others
> to steeel because there is nothing to steal. Of course it blocks on
> blocking functions.
>
> The fjTask and TBB allows to submit on the pool thread or on the internal
> queue (parent task concept), schedules recursively when blocking and allows
> others working thread to steal tasks from its internal queue.

My opinion:

1) work-stealing will be used when possible (can not disabled)
2) fork/join: recursivly executing arbitary task from the pool is wrong - only 
tasks-trees should be executed recursivly
3) forthe problem a sub-taks waits for an notication (condition variable) from 
its paren-task can be solved with fibers

item 1) is implemented
item 2) and 3) I#ve think about intreface and implementation

regards,
Oliver