Weird issue - threads bottleneck in post-proxy section

[Alan DeKok]

On Mar 2, 2015, at 11:02 AM, RESTOUX, Loïc <loic.restoux at capgemini.com> wrote:
> Yes. We wrote a simple module, which sleeps a little every time it is invoked (see attached file). We replaced rlm_rest with it and launched the benchmark tests again, with the exact same results:
> - pre-proxy: 1250 r/s, all threads are active.
> - post-proxy: 42 r/s, only part of the available threads are active.

  That’s bad.

> I'm wondering if there is a critical section in proxy replies handling which may explain such results ?

  No.  There’s code which handles the post-proxy packet.  But that does the same thing whether or not there are any modules in post-proxy.  And the code dealing with modules in post-proxy is *exactly* the same for post-proxy, authorize, authenticate, etc.

  Ive managed to do 50K PPS proxying with the standard config.  So it makes me wonder what else is going on.

  What happens when you use the “ok” module instead of your custom one?  It might be down to bad interactions with the scheduler.

  Alan DeKok.