clock_gettime_mono in the VFS.
rb at sernet.de
Fri Mar 4 20:52:35 UTC 2016
On Fri, Mar 04, 2016 at 11:28:15AM -0800, Jeremy Allison wrote:
> On Fri, Mar 04, 2016 at 01:18:20PM -0500, Ira Cooper wrote:
> > I'm sorry that I'm coming in late on this one. I've been a bit head down
> > recently.
> > On some platforms clock_gettime_mono has a real measurable cost, alas. I
> > discovered this on Solaris/Illumos when I compiled in the timing code in
> > the samba 3.6 profiling framework.
> Just to follow up, I was OK with this in the review as at least on x86-64-linux, the
> gettime calls aren't even real system calls - just a memcpy. It's
> different on ARM I know.
> > I can understand why you'd want this on Linux, getting this data is pretty
> > painful, especially on older kernels.
> > But it should be a compile time option, so we don't impose a cost on
> > platforms where it has real cost, and on people who don't want it.
> > Hopefully this is something we can add. :)
> > Note: I expect over-time, this type of profiling will go down in
> > usefulness. Linux tools to do this type of measurement are improving
> > quickly, and dtrace can do it fairly well on other platforms.
> > But today.... I totally get it, I just don't want us to regress on other
> > platforms.
> Fair enough - if you want to add something that allows this
> as a configure-time option I'm happy to review. I'd suggest
> it's on by default - that allows vendors who really
> don't want it to disable. Ralph, do you agree ?
Not sure, but one way to do it could be to make it part of
--with-profiling-data, which would mean it'd be disabled by
default. That would do it for my use case.
Fwiw, it seems on Solaris one would use gethrtime for this kind of
SerNet GmbH, Bahnhofsallee 1b, 37081 Göttingen
phone: +49-551-370000-0, fax: +49-551-370000-9
AG Göttingen, HRB 2816, GF: Dr. Johannes Loxen
More information about the samba-technical