On Wed, 31 Dec 2003, Manuel Kasper wrote:
> Fred Wright said:
> > I presume the HZ increase was only in that version, since the current
> > version seems to have 250. Cranking it up to 1000 is actually a pretty
> > bad idea - see below.
> Yes, it was only for polling. From sys/i386/conf/LINT:
> # It is strongly recommended to use HZ=1000 or 2000 with DEVICE_POLLING
> # to achieve smoother behaviour.
2000? Yikes! Well, I suppose the clock service stays in the cache, with
only a few cycles of user code per tick to chase anything out. :-)
> > Only if wireless cards can generate sufficiently high packet rates to
> > cause trouble, which may well not be the case.
> Maybe not exactly "high packet rates", but apparently they can still
> generate enough CPU load, especially with WEP and/or IPsec...
Doing such complex processing at interrupt level isn't a very good idea,
but BSD (and Linux as well, I'll bet) doesn't make it easy to do it
efficiently any other way.
> > (well, OK, 500us average) is pretty significant. What's really wanted is
> > a setup where packet processing is only delayed when the system is
> > "getting in trouble". And of course at some point you have to start
> > throwing away packets when you can't keep up.
> Yep, but let's leave that to the FreeBSD developers, shall we? ;)
Sure. Those were meant as general comments.