Hi Dan,
Thanks for the contributions, I will review
https://github.com/OpenSIPS/opensips/commit/04175e5728b3ab2586502627f54aaab2db440999
And the docs you are looking for is
https://www.opensips.org/Documentation/Script-CoreParameters-3-0, right ?
The "inconsistency" you mentioned has deeper roots and comes from how
UDP and TCP are actually implemented in OpenSIPS (like for UDP each
interface has its own set of processes, while for TCP there is only one
shared pool of procs). The auto-scaling simply followed that.
Best regards,
Bogdan-Andrei Iancu
OpenSIPS Founder and Developer
https://www.opensips-solutions.com
OpenSIPS Summit 2019
https://www.opensips.org/events/Summit-2019Amsterdam/
On 04/30/2019 01:00 PM, Dan Pascu wrote:
On 26 Apr 2019, at 12:00, Bogdan-Andrei Iancu wrote:
Well, at the time I developed the feature, that was the feeling in do it -
probably the idea of the difference between UDP and TCP is too deep in my brain.
On the other hand, it make sense what you also say. If you open a bug report
for this, we might be able to pack it as fix for 3.0.
I pushed a small fix that allows one to specify an auto scaling profile with
udp_workers. Someone more knowledgeable with the code should double check it,
just in case I missed something.
It works by defining a global udp scaling profile and uses that if the listener
interface didn't define any (i.e. it acts as a fallback:
if listener_scaling_defined:
use_listener_scaler
elif global_scaling_defined:
use_global_scaler
else:
use_none
This means that if both listen and udp_workers define a scaling profile, the
one defined in listen will take precedence.
Someone should update the documentation with this detail (btw, where is the
documentation for the global script parameters? I checked the source tree and
could not find it)
As a note, there is some inconsistency in the behavior compared to tcp. With
tcp, since it can only define the scaling profile with tcp_workers, that
profile is global and is applied to all listening interfaces. So if one has
multiple listeners for tcp, all will share the same single pool of workers.
With udp though, because it checks it per listener and it falls back to the
global one if undefined, there will be a worker pool per listener interface,
even if the listener doesn't define it's own scaling profile. So because of the
implementation details, udp will have as many worker pools as udp listening
interfaces are defined.
In an ideal scenario, the scaling defined globally with udp_workers and
tcp_workers should only create one worker pool shared by all interfaces that
don't define their own, and only interfaces that explicitly define a scaling
profile should have their own separate worker pool. However the changes
required to do this are not trivial and seem to require a major refactoring of
the code so I left them out for now. Maybe in a future update someone more
knowledgeable could reorganize the code to achieve this and also allow defining
a scaling profile with tcp listeners.
On 04/25/2019 08:31 PM, Dan Pascu wrote:
On 25 Apr 2019, at 19:32, Bogdan-Andrei Iancu wrote:
Dan,
It seems to be a documentation error. The udp_workers param do not support the
"use_auto_scaling" option, as for UDP you can define the profile per listener.
I will fix the doc.
It would make sense to be able to specify them in a similar manner. Either
through xxx_worker or listen or both.
I'd like to keep my configuration organized. Right now I can only specify it
through tcp_workers for tcp and listen for udp. Why is there this limitation
that udp_workers doesn't accept the syntax and listen only accepts it for udp?
--
Dan
_______________________________________________
Devel mailing list
[email protected]
http://lists.opensips.org/cgi-bin/mailman/listinfo/devel
