Re: [Xenomai-core] Analogy DIO speed

[Alexis Berlemont <berlemont.hauw@domain.hid>]

Hi,

Stefan Schaal wrote:
> Thanks to Alexis's Analogy development, digital I/O is possible with Xenomai using National Instrument DAQs. While the basic DIO functionality works in the most current xenomai-head, I am  wondering how to achieve maximal I/O speed with the 32 bit digital I/O sub-device on my NI6259 card. 
> 
> The options are
> 
> 1) single acquisition with a4l_sync_dio()
> 2) instructions and instruction lists using a4l_snd_insnlist() and a4l_snd_insn()
> 3) streaming acquisition with commands (e.g., a4l_snd_command() and related functions)
> 
> My question concerns how many 32-bit DIO instructions per second I
> should be able to achieve with the various options.
> 
> For instance, option 1) seems to take about 5000 nanoseconds on my
> Ubuntu 8 core i386 computer (3Ghz processors).
> Is this normal? Or should it be faster?
> 
> Option 2) seems to give me about 50% speed up, i.e., roughly 3500
> nanoseconds per DIO.

With a4l_sync_dio and/or instructions, you perform one ioctl for each
acquisition.

With instruction lists, you perform less syscalls but the count of
copies (user <-> kernel space) is the same.

So 3.5 µs for:
- switching from user to kernel space
- copying a little structure from user space
- calling the suitable insn_bits handler
- performing PCI I/O
- copying data back (few bytes) to user space
- switching back to user space
It is not that bad, no ? Does anybody have an accurate idea on the
duration of a common ioctl on such a powerful machine ?

Wasting 1.5 more µs in a4l_sync_dio() is annoying. You already have
noticed that a4l_sync_dio() is a wrapper function which
relies on the instruction ioctl. This function does not make any more
ioctl, it just computes some pointers.

> 
> Would option 3) give me a massive speed up and get me closer to the
> 20Mhz processing power of the NI 6259 board? 
Yes. Instead of making one syscall per acquisition, you would get a far
more sensible ratio. With a4l_mmap, you would even be able to save
copies from kernel space: instead of using a4l_async_read, your user
process would get direct access on the kernel buffer where PCI DMA shots
arrive.

I have some problems with
> implementing commands on my NI6259 so far.
Could you remind me what was the problem ?

> 
> Thanks a lot for any help!
> 
> -Stefan
> 
> _______________________________________________
> Xenomai-core mailing list
> Xenomai-core@domain.hid
> https://mail.gna.org/listinfo/xenomai-core
> 
Alexis.