Re: Scalability enhancements for gettimeofday

[David Mosberger <davidm@napali.hpl.hp.com>]

>>>>> On Thu, 20 May 2004 12:05:57 -0700, Christoph Lameter <clameter@sgi.com> said:

  Christoph> The Altix systems never use ITC to determine the time so
  Christoph> this check and the cmpxchg is not necessary.

I see the problem: the basic problem seems to come from the fact that
the non-ITC-interpolation code didn't quite come out the way I
anticipated: I expected even this code to use/need last_nsec_offset
just like the ITC-based code, but it ended up not doing that.  I
_think_ that's OK and, if so, we should probably push all the
last_nsec_offset back down into the interpolator-specific code.
However, this is tricky code so I want to mull it some more to make
sure I'm not missing a subtle race.

  Christoph> Also the fast system call handler for sys_gettimeofday()
  Christoph> can only improve scalability if the ITC is the basis for
  Christoph> gettimeofday which is also never the case on Altix SN2
  Christoph> systems.

Does this really make a substantial difference?  My guess is no and,
if so, I'd rather not add any more #ifdefs.

	--david