On Wed, Nov 04, 2009 at 02:46:02AM -0500, Thomas Wihl wrote:
> When I try a
> ping 127.0.0.1
> on a Jade with QNX 6.4.0 it takes 3ms, every other system (including QNX on a Power PC) takes <1ms.
> A ping from PPC to Jade takes 1ms, from Jade to PPC 3ms.
> 
> Attahed a kernel log of the ping to localhost.
> Setting the priority of ping very high and killing qconn doesn't change those numbers.
> 
> Any ideas?

Replying to a ping is a lot easier than originating one.
The former receives a packet which goes partway up the
stack, a few fields get flipped and then the same packet
goes back out.  At the os level this amounts to an
interrupt, scheduling and running some code.  The latter
brings in various messages passes which you can see from the
trace are significant on this platform when compared to
these periods.  Ping itself takes a longer path once a
second which is also the default interval.  If this longer
path is taken the extra steps are included in the rtt
calculation.  In the trace it's the extra _IO_MSG,
_IO_WRITE, _IO_MSG (setsockopt(), sendto() and
setsockopt()).  The real meat is the following _IO_WRITE,
_IO_NOTIFY, _IO_READ (send(), poll(), recv()).  To hit the
fast path more often which skips the first three messages
try something like 'ping -i0.5 ...'.  You should see some
1ms, 2ms cases along with the 3ms from the long path on this
platform.  I'll look into excluding the extra long path
steps from the rtt calculation.

-seanb