Re: Wanting opinions->uController OS (kernel re-entry)

["Bryan J. Smith <b.j.smith@ieee.org>" <thebs413@earthlink.net>]

From: "Bryan J. Smith <b.j.smith@ieee.org>" <thebs413@earthlink.net>
> Yes, RTAI reduces a lot of efforts.  But I would still argue RT/Linux
> is the way short of developing a fully blown, re-entrant kernel
> like VxWorks or even "heavier" re-entrant kernels like Irix or Solaris.
> Linux's design cannot scale, let alone for real-time, unlike those
> other alternatives.

FYI, just so everyone knows the _unbiased_ world I'm coming from.

I consider Linus no less than _brilliant_ when he first introduced the
concept of 1 entry per CPU in kernel 2.4.  It is the best balance of 
responsiveness and throughput, minimizing context switching.

Most re-entry kernels either and greatly reduce throughput for
response time -- especially in message passing microkernels like
HURD and, even worse, QNX -- or they introduce massive complexity
which Irix and Solaris do on the "general" end or VxWorks more for
the custom-built/embedded end.

Now there have been hacks that inefficiently utilize the 1 entry per
CPU like in the pre-empt patch.  And there continues to be additional
hacks and redesigns to make Linux more re-entrant.  But IMHO, I
think the future of multi-cores and, even more so, virtual cores should
push everyone to go back to the 1 entry per CPU, now core, approach.

If I need more entry for hard real-time, I'm going to pre-empt the
generic kernel itself.  I'm totally in agreeance with FSMLabs' original
concept -- you don't need to build an entire, re-entrant, real-time
kernel just for a couple of pieces of hardware and/or programs to
run in hard real-time.  You can run a generic kernel as a non-real-time
process atop of a pre-empting real-time kernel with its real-time
processes.

Which is why I hope Linux always stays a "soft" real-time OS in its
generic form.  It's overkill to change it, and doing otherwise will
only affect a lot of generic development which is still the
overwhelming majority of uses for Linux.  And it's still a crapload
better than CE or NT kernels in this regard (50ms response time --
man, that's laughable).



--
Bryan J. Smith   mailto:b.j.smith@ieee.org


-
To unsubscribe, send email to majordomo@silug.org with
"unsubscribe silug-discuss" in the body.