TDMA / Interrupts / Pre-emptible

[Len Gross]

On Dec 7, 2007 2:46 AM, Harti Brandt <hartmut.brandt at dlr.de> wrote:

> On Fri, 7 Dec 2007, Bruce M. Simpson wrote:
>
> BMS>I can't comment on kernel scheduler jitter though, so someone who is
> working
> BMS>directly in that area will hopefully respond -- arch@ or hackers@
> might be a
> BMS>better place to field that question.
> BMS>
> BMS>I believe microsecond resolution for your app should be possible in
> the
> BMS>kernel. If it isn't, I'd like to know why. [It would be really, really
> nice
> BMS>to have better real-time support in FreeBSD, i.e. a deadline
> scheduler.]
>
> A couple of years I did exactly the same as the OP - implementing a
> satellite MAC layer (FM-TDMA) on a cluster of 5 FreeBSD machines. I think
> it was the time when we moved from 4.X to 5.0 or 5.1. I could not get it
> reliable because the jitter in the kernel (I implemented everything in
> netgraph over ethernet) was in the order of several 100 usec. I had HZ at
> 10000 (all my simulation machines run on that). I tried really hard to
> find out where these jitters came from, but failed. Trying to trace the
> timing of execution changed the figures completely each time. Finally I
> gave up, because the project luckily ended :-/ At one point I tried to
> bump HZ to 20000 or so, but at this point TCP broke. I think this might
> have come from the RTT computation which is/was done in ticks and the
> square of something would overflow the variable.
>
> One must also carefully choose the ethernet adapter for this kind of
> things, because it may add any kind of jitter/delay. At that time the best
> where the DEC/intel if_dc types.
>
> Of course with the actual kernel the situation may be quite different.
> Don't know what influence have interrupt threads and this stuff. I would
> be interested to hear how things work out and, yeah, better real-time
> support would be great :-)
>
> Regards,
> harti
>
>

I think the general answer is probably that drivers and Kernel code have the
same lack of determinism as user-land code.  I found a LINUX article
"A measurement- based Analysis of the real-time performance of LINUX" that
seems to support this view.

If I can get latency/jitter of 100 microseconds that would be good enough
for my next step.  Will keep you posted.

-- Len