[asterisk-dev] [Code Review] Hints and devices from hints moved to ao2_container

[sst at sil.at]

-----------------------------------------------------------
This is an automatically generated e-mail. To reply, visit:
https://reviewboard.asterisk.org/r/1003/
-----------------------------------------------------------

Review request for Asterisk Developers.


Summary
-------

handle_statechange in pbx.c has a very great impact on call handling performance. i have noticed this several times and i have written a solution which makes handle_statechange handling 61 times faster.

i have used the patch (rev. 293802 for 1.4) from russellb and jeffpeeler to also fix a deadlock problem and their part to move hints to an ao2_container.

after thinking about the best solution IMO iterating through the whole hint list/container to maybe find a hint with a matching device is very expensive way. So i have used a own container for devices which just have a link back to their hint. 
This means there is a hashed cmp for devices which gives back a multiple object ao2_iterator for all hints, this device, which have a state change, is used in. 

these devices are not longer limited to any length so its possible to use a hint for 100 peers at once.

Tilghman Lesher noticed on the dev-list what would happen when a dynamic hint is used. I have tried this with one dynamic hint and subscribe up to 2500 different users on this, several times and it works just fine. If a subscribe for a dynamic hint comes in, a static extension with static appdata is generated which acts like a normal hint so there is no problem on this. 


Diffs
-----

  trunk/include/asterisk.h 294045 
  trunk/include/asterisk/pbx.h 294045 
  trunk/main/asterisk.c 294045 
  trunk/main/pbx.c 294045 

Diff: https://reviewboard.asterisk.org/r/1003/diff


Testing
-------

dialplan reload drops all devices out of the device container and recreate them by their callback list.
core show hint / hints show all hints also dynamic and their watchers


loadtests shows the following result:

orig version: 
handle_statechange COUNTER: 10009 overalltime: 7041 ms 
MAX: 365 CPS

patched version:
handle_statechange COUNTER: 10009 overalltime: 115 ms
MAX: 580 CPS

same setup, same config tested with sipp.


Thanks,

schmidts