Erik E. Fair <fair@clock.org> writes:

>Um, cesium.clock.org (SPARC LX) is running a kernel from end of April. It
>is slightly different from -current of that time in that it has a pmap
>patch that did not get committed to the tree. I've been meaning to update
>that system to -current, which has a different patch to pmap to handle the
>sun4m pmap panics I was getting.


[Expanding on an earlier reply]

A pmax kernel built on June 15 from approx June 6 source on a pmax
converges to about 5 ppm NTP frequency error on a 5000/240 here.
(Both kernels are on ftp.netbsd.org in  pmax  snapshot directories.)

A kernel built from approx June 22 source converges to about 54 ppm
NTP frequency error on the same machine.  I've rebooted both ways and
trie NTP 3.5f.  It's definitely the kernels which cause the
difference. (the machine is in a stable airconditioned lab.)

I'm wondering if this is related to the PDMA overruns which Paul
Goyette and Bill Studenmund repeated as starting `recently'.

I can't see anything in mips-specific the spl*() or interrupt-handling
code in that period which would explain a difference of the magnitude
I'm seeing.  Nothing in the CVS commit logs screams at me, either.

--Jonathan