[time-nuts] Linux time servers

[M. Warner Losh]

In message: <20090515152610.GI2435 at vanheusden.com>
            Folkert van Heusden <folkert at vanheusden.com> writes:
: > : > You might consider switching to FreeBSD for more reasons than just 
: > : > timing, It's much faster than Fedora and I found the new 7.1 version easy 
: > : 
: > : much faster in what respect? tested how?
: > 
: > The usual benchmark that's cited here is the mysql tps scaling better
: > than Linux.  See for example
: > http://people.freebsd.org/~kris/scaling/mysql.html.  The numbers in
: > this "paper" are a little dated (being for 7.0 and a little over a
: > year old), they show good scaling.  Of course, this isn't the place to
: > debate that in extreme detail (for example, there are other pages I
: > can't find right now that show newer versions of Linux, some better,
: > some worse as changes to the scheduler help or hurt performance).  It
: > is no longer the case that you can automatically assume Linux performs
: > better.  You have to measure things and make sure you use the system
: > that best matches your performance requirements.
: > Also, on a 7.1R system, you need to make sure that you enable tagged
: > queueing.  A last minute change botched it.  7.2R is out now too.
: 
: You can't say that freebsd is faster than linux; you specifically need
: to specify what version of freebsd and what version of linux you're
: using. Also the hardware platform matters as well as the compiler (and
: version) used to compile mysql and numerous other parameters.

I didn't say FreeBSD was faster than Linux.  Please read what I said
carefully (note, the quotes stuff at the top isn't me).  I said for
some work loads, it is faster.

: What would be interesting is how a specific linux-kernel with the pps
: patches by rodolpho compare to a specific freebsd version with the same
: ntpd compiled using the same gcc and such.

Yes.  All my evaluations of Linux pre-date this patch.  However, even
with the patch, I can still say that Linux performs worse than FreeBSD
on NTP because the patch hasn't been committed to the kernel.org tree.
I guess this is the difference between "Linux can be made to perform
better with this out-of-tree patch" and "Out of the box, Linux
performs well."

When FreeBSD switched from gcc 3.x to gcc 4.x, I did measurements of
the ability of the kernel to track a PPS (also changes with the major
revision of the kernel).  I found that there was no measurable
difference between the different FreeBSD kernels I tested despite
being built with a number of different compilers (3.4.5, 4.2.0 and
4.1).  It turns out that the algorithms for steering the time aren't
dependent on how fast the results are computed, but rather dependent
on the results being computed correctly.  I will admit that my testing
of Linux was been rather cursory over the years compared to the
attention I've given to FreeBSD.

Of course, we're mixing up problems a little bit here.  The ntpd with
pps performance issue is somewhat different than the claims another
writer was making about FreeBSD being faster for his servers...

Warner