comp.lang.ada
 help / color / mirror / Atom feed
From: Ludovic Brenta <ludovic@ludovic-brenta.org>
Subject: Re: How to determine if task is busy or not?
Date: Sat, 8 Aug 2009 11:39:11 -0700 (PDT)
Date: 2009-08-08T11:39:11-07:00	[thread overview]
Message-ID: <288fa06f-bae4-4e60-8b14-79e4bf5a96d5@z34g2000vbl.googlegroups.com> (raw)
In-Reply-To: 11m01goasspup.x63didgf7t0x.dlg@40tude.net

Dmitry A. Kazakov wrote on comp.lang.ada:
> Maciej Sobczak wrote:
>> Dmitry A. Kazakov wrote:
>
>>> The scheme described by Ludovic is not only simpler, it is also race
>>> condition free. The reverse scheme has a race condition in it. If you have
>>> asked if a task is idle, you do not know if it still is right now
>
>> Depends. If it is the asking entity that is also providing the job,
>> then the problem above cannot happen. In other words, if the "manager"
>> asks the task whether it is busy or idle and the answer is "idle",
>> then that answer it true until the same manager provides new work
>> unit, because there is no other way for the task to switch from "idle"
>> to "busy". Which means that the manager can always safely handle a new
>> work unit to the "idle" task, no race is possible.
>
> As I said, only if single job supplier (manager) is here. In that case it
> is still a poor scheme because of polling. A better (I tempted to say
> proper) one is when idle tasks queue themselves into the pool of idle tasks
> maintained by the manager. But a protected object of tasks would be better
> here as well.
>
>> The advantage of such a setup is that the manager knows whether the
>> work can be handled *immediately* by any one of the tasks in the pool.
>> False negatives can be harmless, as long as the positives are
>> accurate.
>
> No, it is not an advantage.
>
> 1. Job is done to get a result, it is a lengthy process, so there is a
> worker to perform it asynchronously. Until the job is done it is of little
> interest whether it has been started or not. (Some interest exists if jobs
> are cancelable if not initiated)
>
> 2. The manager is only a mediator here. If a job cannot be started
> immediately what the manger is supposed to do? If it does not cancel the
> job, the only option is to wait. This is what the job queue is for.
>
>> Using a queue, as suggested by Ludovic, makes sense only when the work
>> units can *wait* for being processed. This might or might not be the
>> case.
>
> See above. If there is some service timeout a queue is serviced by a
> monitor task (a special "mister no" worker). Still better.
>
>> There are examples of systems where either of these schemes are good
>> (batch processing vs. real-time?). I would not say that one is better
>> than another.
>
>> To Tomek - the simplest way to deploy your original idea is to use a
>> flag (Busy/Idle - an enumeration type is perfect here), which is
>> shared and used by both the worker task and its manager. Protected
>> object is a proper solution here and you might even use the same
>> protected object for managing both the flag and the work unit itself.
>
> No, you don't need any extra shared data. Do conditional entry call do
> determine if the worker is ready:
>
> task body Worker is
> begin
>    loop
>       accept New_Job (Work_Item : in out Job);
>       ... -- Service
>    end loop;
> end Worker;
>
> The manager does a conditional entry call (RM 9.7.3):
>
>    select
>       Worker.New_Job (Thing_To_Do);
>    else
>       -- The chap is busy
>       ...
>    end select;

But that becomes unwieldy in the presence of multiple worker tasks,
any one of which needs to be idle when enqueuing a job.  How about:

protected Queue is
   procedure Enqueue (J : in Job);
   entry Get_Next (J : out Job);
private
   -- left as an exercise :)
end Queue;

protected body Queue is
   procedure Enqueue (J : in Job) is
   begin
      if Get_Next_Job'Count < 1 then
         raise No_Worker_Is_Presently_Waiting_For_A_Job;
      else
         -- ...
      end if;
   end Enqueue;

   entry Get_Next (J : out Job) when Jobs_Enqueued is
   begin
      ...
   end Get_Next;
end Queue;

Start all workers before the manager; all workers then start waiting
on the Get_Next entry.  The manager keeps enqueuing new jobs until all
workers are busy, at which point no worker is waiting on Get_Next
anymore and the manager gets an exception to signal that condition.

Then the difficult part is to ensure that all workers collectively
consume jobs at least as fast as the manager can enqueue them.

--
Ludovic Brenta.



      reply	other threads:[~2009-08-08 18:39 UTC|newest]

Thread overview: 7+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2009-08-08 10:49 How to determine if task is busy or not? Tomek Walkuski
2009-08-08 10:56 ` Ludovic Brenta
2009-08-08 11:27   ` Dmitry A. Kazakov
2009-08-08 14:59     ` Maciej Sobczak
2009-08-08 16:00       ` Tomek Wałkuski
2009-08-08 18:22       ` Dmitry A. Kazakov
2009-08-08 18:39         ` Ludovic Brenta [this message]
replies disabled

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox