[prev in list] [next in list] [prev in thread] [next in thread] 

List:       sr-dev
Subject:    [Devel] core timer and TM retransmission timers
From:       jf-smig () ibelgique ! com (=?ISO-8859-1?Q?Jean-Fran=E7ois?= SMIGIELSKI)
Date:       2006-12-27 10:46:51
Message-ID: 20061227094919.D4ADB48667 () ifrance00 ! ifrance ! com
[Download RAW message or body]

The patch affects a bit the performance.

I tested the new openser with a SIPp that sent message that where statefully \
forwarded and voluntarily not answered (so, message stay as long as possible in \
memory). Then I compared the results with those about a regular 1.1.0 openser.

I started 300 new transactions/second (the limit came from the shared memory limited \
to 128MB, and with the size of the messages I sent it maid quite 20000 transactions \
in memory). I observed an increase of the CPU of about 25%. To reduce this overhead, \
I decided to loose precision by snapping wake up time on a "5ms grid". So, I could \
wake up less often for more work. It worked and reduced the CPU usage, but not below \
the threshold of the regular OpenSER.

I made this test on a weak machine, and real benchmarks are surely necessary.

I prepare the patch and I post it as soon as possible.

Thanks!


Klaus Darilion:
> Hi Jean-François!
> 
> Please post your patch on the bug tracker on sourceforge:
> http://sourceforge.net/tracker/?group_id=139143&atid=743022
> 
> thanks for your work.
> 
> Do you know how your patch affects the timer performance of openser?
> 
> regards
> klaus
> 
> 
> 
> > Hello again,
> > 
> > I proposed several weeks ago to change the core timer in OpenSER to get more \
> > precision. This work is now done and my openser seems to work fine.
> > It was achieved with precise 'select' sleeps and the expirations are now \
> > evaluated on timeval structures. The previous "jiffies" timer now registers on \
> > the "select" timer. 
> > I also proposed to use this precise timer to manage the TM retransmission timers.
> > This is also done and seems to work fine. TM now conforms to the RFC3261 (with a \
> > different INVITE retransmission management). This was done with a direct \
> > registration of TM on the precise timer (no more on the jiffies timer), and with \
> > timer lists only for INVITE. 
> > But I had to do it first for openser 1.1.0, and I am preparing a port on the cvs \
> > sources. There is no real problem with that, I wrote new files and changed only a \
> > few. 
> > I would submit my evolution to your analysis, but I don't know how.
> > Can I post a patch on this list?
> > 
> > JF Smigielski
> > 
> > 
> > ________________________________________________________________________
> > iBELGIQUE, exprimez-vous !
> > http://web.ibelgique.com/
> > 
> > 
> > ------------------------------------------------------------------------
> > 
> > _______________________________________________
> > Devel mailing list
> > Devel@openser.org
> > http://openser.org/cgi-bin/mailman/listinfo/devel
> 
> 
> -- 
> Klaus Darilion
> nic.at
> 
> 


________________________________________________________________________
iBELGIQUE, exprimez-vous !
http://web.ibelgique.com/


[prev in list] [next in list] [prev in thread] [next in thread]