Re: sleep in microsecond or nanosecond

[olafBuddenhagen]

Hi,

On Mon, Aug 16, 2010 at 05:07:02PM +0200, Samuel Thibault wrote:
> olafBuddenhagen@gmx.net, le Fri 13 Aug 2010 08:42:33 +0200, a écrit :

> > I think improving the timeout granularity in general would be rather
> > complicated, and make little sense...
> 
> Improving the timeout granularity for even just one case is most of
> the work: programming e.g. the HPET timer, managing wake-ups, etc.
> Plugging that to the usual kernel timeout functions should really be
> trivial after that.

My point is mostly that for very short dalays, it's probably useful to
do busy-waiting; while for ordinary message calls, it makes no sense.
The overhead of normal message calls would also be problematic for
precise timing.

Note that the message timeouts are always handled by the kernel; so
changing it to a special trap wouldn't cost any flexibility.

Also, using precise timers for all message timeouts would probably add
unnecessary overhead to the normal IPC path.

It seems to me that (mis)using the message call for timing is probably
not such a good idea -- the requirements are quite different.

-antrik-