Re: [realtime] Re: [ANNOUNCE] Adaptive Domain Environment for OperatingSystems

[Mark_H_Johnson@Raytheon.com]

Hmm. I'm somewhat concerned about an apparent attitude of...
 - if the hardware does "X" (e.g., PCI stalls), the software can't do
anything about it
Yet the example of X and "non spec" cards appears to be part of a problem
that a real time developer needs to...
 - know about and understand
 - the software COULD do something different that didn't cause the problem
to begin with
As a system designer, I want to know where the "oopses" are and avoid them.
In our case, we weren't planning on running X anyway, but we might have
similar problems with network I/O, the cluster interconnect, and some
occasional disk I/O. Are there some things we should look for and avoid for
those cases?

We're looking at a cluster of machines to run a real time simulation
(fastest rate 80hz). Its a large system that we are unwilling to run in
kernel mode. Our system needs relatively low latency and jitter. However,
for most of our code, a millisecond here or there isn't a big deal. For a
couple specific hardware interfaces - we have an exceptionally tight
tolerance (+/- 50 microsecond jitter on data transfer, 80hz rate). We'll
use a hardware assist for now - perhaps a kernel RT task later.

The solution we're looking at today is a combination of...
 - low latency patches to Linux
 - some form of "user land" RT for most of our tasks (RT_FIFO, LXRT?)
 - a few items [TBD] to be under RTLinux, ADEOS, or similar product
>From what I've read in the ADEOS docs doesn't help much over what RTLinux
does today other than avoid the patent [which does have value to us]. What
we'd prefer to see more focus on good performance in user mode and still
have access to non-blocking Linux services. I see most of the solutions
suggested to date are unable to meet that need.

--Mark H Johnson
  <mailto:Mark_H_Johnson@raytheon.com>


                                                                                                                             
                    Karim Yaghmour                                                                                           
                    <karym@opersys.com>           To:     Cort Dougan <cort@fsmlabs.com>                                     
                    Sent by:                      cc:     William Montgomery <william@opinicus.com>,                         
                    owner-realtime@realtim        realtime@realtimelinux.org, rtl@rtlinux.org                                
                    elinux.org                    Subject:     Re: [realtime] Re: [ANNOUNCE] Adaptive Domain Environment for 
                                                  Operating Systems                                                          
                                                                                                                             
                    02/16/01 04:51 PM                                                                                        
                                                                                                                             
                                                                                                                             




Hello Cort,

You're somewhat right. There are hardware issues that may be out
of the reach of any software, the PCI stall example is a good
one. On the other, whether RTLinux is running on Alpha, MIPS or
PowerPC, device drivers can still issue the equivalent cli/sti
and ruin determinism. There's nothing RTLinux can do about it.

Best regards,

Karim

Cort Dougan wrote:
> [snip]
> Current versions of X (4.x) are pretty aggressive with cards that are not
> entirely spec which results in PCI stalls.  That's a problem for any
> real-time system since software can't do much about this once it's
> happened.