SV: [summary] HTTP/1.1 client requests causing CPU cycle bug/problem

The problem of heavy CPU-usage still occurs even with HTTP/1.1 handling
turned off, though in a different manner than described by Stipe. Multiple
threads sending lots of simultaneous http-requests to Kannel can have the
exact same effect of sending the boxes in a powerfrenzy.

I remember having seen one of our programmers testing the sendsms interface
with some very early java code which had some (simple) problems with closing
a http-transaction. The code would do an initial first http-request to
Kannel which wasn't closed properly for some reason, and the second the code
did http-request no. 2 smsbox would eat all of CPU. This we tried several
times - always with the same result! (note that this was AFTER I turned 1.1
handling off). It's too long ago for me to show any log files I'm afraid.

There seems to be a problem with Kannel handling multiple http-requests at
the same time. We always have to restart Kannel at least a couple of times a
week because of smsbox gone crazy, and it always occurs after periods of
heavy traffic i.e. after lots and lots of http-requests having been sent to
Kannel.

BTW we're running redhat 6.2.

cheers,

Frederik Ammitzbøll
Unwire
Vestergade 12A, 3.
1456 København K

Tlf.:  +45 33 33 08 70
Mobil: +45 27 11 99 99
Fax :  +45 33 33 09 70
Web:   www.unwire.dk


> Inspired by Jörgs repost of the orginal problem reported by:
>
>   Frederik Ammitzbøll <[email protected]>
>   Tue, 17 Jul 2001 12:07:02 +0200
>   Msg-ID: <[email protected]>
>
> I did some in-deep investigations on this and here is a draft of the
> current summary:
>
> As Frederik pointed out there is a *huge* impact when HTTP interfaces
> are used by HTTP/1.1 compliant user agents (especialy IE). When *any*
> HTTP server running as a thread in *any* box is triggered using a
> HTTP/1.1 request by IE the request gets served and ~1 minute later the
> box runs up to 99% of CPU cycles.
>
> Keep in mind that this has impact on *any* box, which means:
>
>   bearerbox: for admin HTTP interface
>   smsbox: for sendsms/sendota HTTP interface
>   wapbox: for PPG interface
>
> due that all incorporate Kannel's HTTP server implementation from
> gwlib/http.c.
>
> Here is a log to reproduce the effects seen:
>
>   $ cvs update gateway
>   $ cd gateway
>   $ ./configure --with-defaults=speed
>   $ make progs
>   $ cd gw
>   $ ./bearerbox wapkannel.conf
>
> now at least bearerbox is running and listening to port 13000 for
> administration commands.
>
> Now use URL http://<hostname>:13000/ to access the bearerbox from your
> MS Internet Explorer. (Unfortunatly I could not reproduce the effect
> using HTTP/1.1 requests send using telnet by hand, which is indicating
> that IE produces some kind of effect here!)
>
> After ~1 minute you get the following effects on the platforms:
>
> i686-pc-linux-gnu (SuSE 7.x):
> 2002-01-26 02:39:45 [4] DEBUG: HTTP: Resetting HTTPClient for
> '1.2.3.4'.
> (~1 minute passes)
> (CPU cycle rate for bearerbox hits 99%)
>
> sparc-sun-solaris2.6 (E250):
> 2002-01-26 02:00:13 [3] DEBUG: HTTP: Resetting HTTPClient for
> '1.2.3.4'.
> 2002-01-26 02:01:13 [1] ERROR: Error reading from fd 20:
> 2002-01-26 02:01:13 [1] ERROR: System error 131: Connection reset by
> peer
> 2002-01-26 02:01:13 [1] DEBUG: HTTP: Destroying HTTPClient area
> 16f448.
> 2002-01-26 02:01:13 [1] DEBUG: HTTP: Destroying HTTPClient for
> '1.2.3.4'.
> (CPU cylce rate for bearerbox stays normal)
>
> i686-pc-cygwin (Cygwin 1.3.9):
> 2002-01-26 02:39:10 [3] DEBUG: HTTP: Resetting HTTPClient for
> '1.2.3.4'.
> 2002-01-26 02:40:10 [1] ERROR: Error reading from fd 39:
> 2002-01-26 02:40:10 [1] ERROR: System error 104: Connection reset by
> peer
> 2002-01-26 02:40:10 [1] ERROR: Error reading from fd 39:
> 2002-01-26 02:40:10 [1] ERROR: System error 104: Connection reset by
> peer
> 2002-01-26 02:40:10 [1] DEBUG: HTTP: Destroying HTTPClient area
> 0x100dacc0.
> 2002-01-26 02:40:10 [1] DEBUG: HTTP: Destroying HTTPClient for
> '1.2.3.4'.
> (CPU cycle rate for bearerbox stays normal)
>
>
> Note the 1 minute (!) difference in the Solaris *and* Cygwin log
> between the resetting of the HTTPClient and the error!
>
> So we have a different behaviour for the same situation on Linux and
> Solaris. Linux does not give any error message, but the thread seems
> to burn the CPU time. Solaris and Cygwin instead screems but leaves
> the CPU ticks alone.
>
> I guess we have some problem with TCP communication from IE to the
> HTTP servers of Kannel here?!
>
> Any other testers for MacOS X and FreeBSD, who may run this scenario
> and report what their OS does?
>
>
> Stipe
>
> [email protected]
> -------------------------------------------------------------------
> Wapme Systems AG
>
> Münsterstr. 248
> 40470 Düsseldorf
>
> Tel: +49-211-74845-0
> Fax: +49-211-74845-299
>
> E-Mail: [email protected]
> Internet: http://www.wapme-systems.de
> -------------------------------------------------------------------
> wapme.net - wherever you are
>
>