[CONFIG_PREEMPT_RT] High frequency periodic timer

[John Sigler <linux.kernel@free.fr>]

Hello everyone,

[ Please include my address in the CC list ]

I have been experimenting with the CONFIG_PREEMPT_RT patch for a few
months. Specifically kernel 2.6.20.7-rt8.

http://rt.wiki.kernel.org/index.php/Main_Page

I've been running into some unexpected problems, so I wanted to ask
those who have some experience with this patch what they thought.

Here's a brief description of my application:

System A periodically sends timestamped packets. These packets travel
over a network, or through a network emulator, and get randomly dropped
or delayed by a random amount of time. These packets reach system B at
some point. System B uses the timestamp to re-send the packets smoothly,
as if they had been produced on B, with only a constant time-shift.
Basically, B acts as a "dejittering" router.

I have implemented this with high-resolution timers. Every time the
timer fires, I send one packet. In "parallel", I buffer incoming packets
that arrive from system A.

Consider 1316-byte packets and a 80 Mbit/s stream bit rate.

This means the timer period is 1316*8 / 80 = 131.6 µs

I've been wondering whether having such a small period (high frequency)
might be a problem / challenge for the real-time kernel (scheduler, IRQ
handler, other component).

The CPUs used are mid-range (i.e. 1.3 GHz P3 and 1.5 GHz Celeron M) 
single core, and they easily copes with just the receiving part.

But I have seen some odd behavior (random crashes in my program) that
must mean I have made some incorrect assumptions.

What do you all think?

Regards.