Garbage collection of completed tasks

[danichcr72]

Hi,

Another general issue that appeared in my experience with fiber is that completed tasks do not release their memory until the dispatcher/system fiber runs.

The dispatcher fiber is not scheduled in all (?) / most(?) schedulers unless there are no application fibers pending. It would be great to provide another mechanism to garbage collect completed tasks, either by the user calling a function or by having the dispatcher fiber scheduled with some other priority.

Thanks,

Daniel Chavarria
Trovares Inc.

[olk]

One of the tasks of the dispatcher-fiber is to release terminated worker-fibers (== fibers created by the user-code). The dispatcher-fiber is the first fiber in the ready-queue when the application/thread starts. It always enqueues itself before it hands-over to another worker-fiber (dequeued from the ready-queue).
So the dispatcher-fiber should be executed multiple times in your app.

[danichcr72]

Thanks for the quick reply. I can see that the dispatcher fiber is scheduled to run, however in the shared_work scheduler at least, it does not run unless there are no worker fibers pending:

if ( ! rqueue_.empty() ) { /*<
        pop an item from the ready queue
    >*/
    ctx = rqueue_.front();
    rqueue_.pop_front();
    lk.unlock();
    BOOST_ASSERT( nullptr != ctx);
    context::active()->attach( ctx); /*<
        attach context to current scheduler via the active fiber
        of this thread
    >*/
} else {
    lk.unlock();
    if ( ! lqueue_.empty() ) { /*<
            nothing in the ready queue, return main or dispatcher fiber
        >*/

[olk]

right - shared_work was previously in the examples , so I missed the implementation details

[olk]

releated to #112

[olk]

The problem is that the dispatcher-fiber is not correctly scheduled in the shared_work algo.
If you take a look at work_stealgin algo (branch develop) you see that the dipatcher-fiber is scheduled like in round_robin.
I suggest to switch to work_stealing because it is faster (locality of data) than shared_work.
Otherwise shared_work needs some counters that manage teh scheduling of main-/dispatcher-fiber ...

[olk]

use work-stealing scheduler instead