On 14/06/2016 04:09, Robert Haas wrote:
> On Mon, Jun 13, 2016 at 5:42 AM, Julien Rouhaud
> <julien.rouhaud@dalibo.com> wrote:
>> Agreed, and fixed in attached v3.
>
> I don't entirely like the new logic in
> RegisterDynamicBackgroundWorker.
I'm not that happy with it too. We can avoid iterating over every slots
if the feature isn't activated though (max_parallel_workers >=
max_worker_processes).
> I wonder if we can't drive this off
> of a couple of counters, instead of having the process registering the
> background worker iterate over every slot. Suppose we add two
> counters to BackgroundWorkerArray, parallel_register_count and
> parallel_terminate_count. Whenever a backend successfully registers a
> parallel worker, it increments parallel_register_count. Whenever the
> postmaster marks a parallel wokrer slot as no longer in use, it
> increments parallel_terminate_count. Then, the number of active
> parallel workers is just parallel_register_count -
> parallel_terminate_count. (We can't have the postmaster and the
> backends share the same counter, because then it would need locking,
> and the postmaster can't try to take spinlocks - can't even use
> atomics, because those might be emulated using spinlocks.)
>
I wanted to maintain counters at first, but it seemed more invasive, and
I thought that the max_parallel_worker would be ueful in environnements
where there're lots of parallel workers and dynamic workers used, so
finding a free slot would require iterating over most of the slots most
of the time anyway. I'm of course also ok with maintaining counters.
> If we want to allow the number of parallel workers started to be available
> for statistical purposes, we can keep to uint32 values for that
> (parallel_register_count_lo and parallel_register_count_hi, for
> example), and increment the second one whenever the first one rolls
> over to zero.
>
I didn't think about monitoring. I'm not sure if this counter would be
really helpful without also having the number of time a parallel worker
couldn't be launched (and I'd really like to have this one).
--
Julien Rouhaud
http://dalibo.com - http://dalibo.org
