[Xenomai-help] Re: RTAI porting to ppc

[barbalace@domain.hid]

> barbalace@domain.hid wrote:
> > Quoting Jan Kiszka <jan.kiszka@domain.hid>:
> >
> >> I interpret this as "too slow for me", correct? Do you mind posting your
> >> test scenario (board/CPU, ipipe/Xenomai/kernel versions, test program,
> >> etc.) and your requirements to a related list? You will be welcome. Just
> >> make sure that you test under adequate load, because anything else is of
> >> limited use for a hard RT benchmark. See also xenomai/TROUBLESHOOTING.
> >
> > test scenario is MVME5500/MPC7455, ADEOS/ipipe ppc-1.3-05 Xenomai2.2
> > kernel2.6.14, test program consist in a simple real-time interrupt handler
> in
> > kernel space registered with Xenomai/native skin. The interrupt is
> generated
> > from a VME board (with know response time) on an clock signal. The
> interrupt
> > handler simple write a VME location generating a square clock where the
> level
> > of the output change every interrupt.
>
> OK.
>
> Why not Xenomai 2.2.5 with pipe-1.5? Moreover, current development head
> is 2.6.18, soon .19 (Wolfgang is currently preparing the first PowerPC
> tree ports). Just to avoid that some micros got optimised meanwhile.

The BSP for my board is for kernel 2.6.14 ;-) , I've patched with Wolfgang
gt64260 pic code for ipipe, some ipipe patch >2.6.14 write on flash memory and
the board must be send to Motorola. So I prefer to stay to 2.6.14 for the
moment.

> > Is a very simple scenario. I only want to measure the interrupt dispatching
> time
> > to compare between Linux/Xenomai/RTAI/VxWorks. The jitter in scheduling and
> in
> > the interrupt dispatching I compare too. I know that under high load system
> > lost  a deterministic beahavior.
>
> A hard real-time system doesn't lose determinism, but it will surely
> expose worse numbers. Still, even lightly unloaded boxes can show high
> latencies - it just takes much longer to happen.
Ok

> > I made some test with system load and unload.
> > With a loaded system the same test in Linux-kernel on Linux/ipipe Xenomai
> > mounted are very slowly compared to Xenomai registered ones.
>
> And I have to understand what you mean. What are the two scenarios
> precisely that you compare under load here?
>
> > I read the xenomai/TROUBLESHOOTING but latency killers are off in my
> kernel. Now
>
> I was referring to TROUBLESHOOTING for appropriate load like the cache
> calibrator. People often think that ping -f or similar tight loops are
> already triggering the worst case.

A... under load for me is intended with data transfer on (ethernet transefers)
or/and some data intesive calculation (matrix transpose with altivec and
without).
I use the same C code (with they API) for Linux, Xenomai and VxWorks. (For the
moment I don't try VxWorks skin)

I've not think at the cache calibrator issue... This could be a very interesting
thing.

> > the situation is: VxWorks 73uS, Linux pure 79uS and Xenomai 78uS, all
> results
>
> Well, I heavily doubt that those 79 us over vanilla Linux remain stable
> if you let your test run for an hour or longer and keep the system
> loaded. Or are you testing with some -rt patch applied?

Test doesn't remain stable under load neither for one minute, you are right.
I don't use rt patch.

> > all well. This test use a semaphore to start data acquisition on VME, the
> > semaphore is signaled in the interrupt handler. This result are the same if
> the
> > system is loaded for VxWorks and Xenomai but not for Linux, it depends on
> the
> > load the acquisition could be 100uS or more.
>
> IIRC, VxWorks is MMU-less. Did you configure Xenomai without user-space
> support as well (.config would be interesting, also when posting to the
> list)? Of course, this only makes sense of you plan to push everything
> RT-wise into kernel, and this is not recommended (fault confinement,
> debugging, legal issues, ...).

For the moment I'm only in kernel space, but I think to go user-space in
january.
Do you think MMU overhead is about a 1uS?

>
> >
> > Another detail: a read/write operation on the VME bus take between 1 and
> 2uS. I
> > need 32  read and one write. The IRQ chain 'round' take some time... I plan
> to
> > review VME/VITA specification jet to determin a possible required time.
>
> Means that most of your ~80 us could be VME access?

Yes, this could be true. Some test tell me that a read (that required VME bus
access and handshaking) require (1,5 - 2)uS; get_tlb() from the first read to
the first write tell me values in the range [53 - 58]uS (more or less). With
Linux. Some tests with VxWorks are not jet done. Interrupt acknowledge
daisy-chain I think require a couple of uS.

> >
> > With RTAI immediate dispatching I want to see if we can go below the
> VxWorks
> > required time.
>
> Already tried the latency tracer with Xenomai? It can show you roughly
> what latency some part of the IRQ path causes ("roughly", because the
> tracer comes with an overhead and cache disturbances). I will also give
> you a clue what RTAI may improve and what is hardware related (before
> writing a single painful line of code). See, once again, irqbench on how
> to trigger a back-trace on the longest delay.
>
A... Another interesting idea! I plan to try it.

> >
> > Do you think I must change some settings? From my point of view results are
> ok.
>
> I personally cannot comment on your PPC board, but on the general
> picture: I once measured 10% better worst-case with RTAI on low-end x86
> (user-space loop), so you /may/ get below current Xenomai numbers and
> save a few micros. The cons are that you have to port, stabilise, and
> maintain quite some code unless you plan this as a disposable
> development. Something you should look at from both sides. Still,
> design-wise sane optimisations are welcome at any time, and I also
> pushed several patches of this kind into I-pipe and Xenomai over the
> last year.
>
> Jan

>From this point I agree to you that I feel Xenomai community more open to ppc
architecture then RTAI.

Thanks a lot,
Antonio