[coyotos-dev] Sleep, wakeup, and persistence

[Sam Mason]

On Sat, Sep 15, 2007 at 02:30:02PM +0200, Valerio Bellizzomi wrote:
> On 15/09/2007, at 3.38, Sam Mason wrote:
> >The NTP daemon will correct up to 500 PPM (parts per million) of error
> >in your clock by changing the rate at which time is incremented.  As the
> >daemon confidence increases in your clock it'll slow down the polling
> >frequency to (normally) about once every 20 minutes.  You just have to
> >leave it running long enough and it'll figure everything out on its own.
> 
> You mean polling a server? once every 20 minutes seems a high rate to me.

If you're trying to keep millisecond accuracy (i.e. what NTP aims to
do), on stock PC hardware, it's not bad.  If you've got a very nice,
probably temperature compensated, crystal you could probably decrease
the polling rate a little.  For most people, a few seconds either way
aren't going to make much difference so you could poll a thousand times
less frequently.

Bear in mind that if you are indeed polling every 20 minutes that means
that a single server could be polled by 12000 clients and it would only
see 10 packets per second.  I've got a pretty basic box (i.e. VIA 500MHz
processor) serving time and it easily (i.e. much less than 10% CPU
usage) keeps up with 200 requests per second.  The thing that fails is
that I've got a consumer level router that can't handle all these new
UDP "connections".


  Sam