Re: Anyone using Ada with Real-Time Linux?

[Simon Wright <simon.j.wright@amsjv.com>]

dmitry@elros.cbb-automation.de (Dmitry A. Kazakov) writes:

> On Sun, 25 Nov 2001 20:50:12 GMT, "DuckE" <nospam_steved94@home.com>
> wrote:
> 
> >I have an application that acquires data at 1 msec intervals, processes the
> >accumulated data and  transmits a summary of the data across a network every
> >330 msec.  The application is currently written in C and runs on VxWorks.
> >Moving the application to a different target, still under VxWorks means
> >shelling out another $30K to Wind River, plus $200 per target (at least).
> >If I can come up with a less expensive choice using Ada, ths may provide an
> >opportunity for a swithc.
> 
> A question to wise men. Is above achievable under "normal" Linux,
> provided that only Ada application is running and tasks are
> dispatched by Ada RTS? Will hard disk I/O be a problem? TCP/IP
> socket I/O?

The main problem is the 1ms period. Under Linux 2.2 you can set the
kernel jiffy down to 0.5 ms by altering HZ to 2000 in asm/param.h,
alter linux/kernel/sched.c:sys_nanosleep to remove the busy-wait on
short delay requests up to 2 ms, and rebuild the kernel. You really
ought to rebuild the standard library too, since Linux doesn't have a
complete sysconf, but it seems to work pretty well (we've not used
this in production work yet, just in a testbed).

The reason for setting HZ to 2000 is that GNAT uses nanosleep to
implement delay, and nanosleep delays for 1 more jiffy than you
expected so that it can be sure to delay for at least the time
requested.

The final trick is to run your tasking Ada program as root. This puts
it into the SCHED_FIFO real-time dispatching class (to borrow a
Solaris-ism). There is a way to get SCHED_RR instead (under 3.14a1,
anyway; I expect it's one of the configuration pragmas if it's
anything like Solaris).

Whether this would fit the requirement I don't know; there might be
unavoidable delays, as Dmitry says.