HZ=1000 on slow CPUs considered harmful?
Robert Watson
rwatson at FreeBSD.org
Wed Feb 23 14:17:22 GMT 2005
On Wed, 23 Feb 2005, Chris wrote:
> I have noticed issues with HZ=1000 on a celeron 2ghz and a AMD64 3.2ghz,
> both caused problems with glftpd app and I had to recompile kernel back
> to default HZ=100 to fix, I have also noticed HZ=1000 add latency on my
> celeron box.
FYI, the recent tcp_subr.c changes I merged may help with performance at
higher HZ levels. Basically, the tcp_isn_tick() function, which stirs up
the initial sequence number pot at intervals, was running ten times as
often as it needed to, as it ran each tick rather than at a fixed
rate/second.
Robert N M Watson
>
> Chris
>
>
> On Tue, 22 Feb 2005 14:00:09 -0800, Kevin Oberman <oberman at es.net> wrote:
> > > Date: Tue, 22 Feb 2005 22:48:21 +0100
> > > From: cpghost at cordula.ws
> > >
> > > On Tue, Feb 22, 2005 at 07:56:03PM +0000, Robert Watson wrote:
> > > > In 6-CURRENT, HZ is 1000 for amd64, i386, and ia64, but 100 for other
> > > > platforms (i.e., ppc, arm, and alpha). I'm not opposed to merging the HZ
> > > > change to RELENG_5 at some point, but given that occasional nits, such as
> > > > the TCP nit, are turning up, I think it's worth waiting until after 5.4.
> > >
> > > Wouldn't that be a problem for slow CPUs like VIA C3 (EPIA) or GEODE
> > > (Soekris)? For fast CPUs, it's not that much overhead, but for slow
> > > CPUs?
> > >
> > > Can HZ remain user-configurable?
> > >
> > > > Robert N M Watson
> > >
> > > Thanks,
> > > -cpghost.
> > >
> > > --
> > > Cordula's Web. http://www.cordula.ws/
> > >
> >
> > As far as I know, no one is talking about removing the knob. I believe
> > the issue is the default value changing from 100 to 1000. There are many
> > boxes that would not be very happy with 1000. I'm sure that there are
> > many 75 MHz Pentiums and 66 MHz 486s still running FreeBSD and even running
> > V5.
> > --
> > R. Kevin Oberman, Network Engineer
> > Energy Sciences Network (ESnet)
> > Ernest O. Lawrence Berkeley National Laboratory (Berkeley Lab)
> > E-mail: oberman at es.net Phone: +1 510 486-8634
> > _______________________________________________
> > freebsd-stable at freebsd.org mailing list
> > http://lists.freebsd.org/mailman/listinfo/freebsd-stable
> > To unsubscribe, send any mail to "freebsd-stable-unsubscribe at freebsd.org"
> >
>
More information about the freebsd-stable
mailing list