[time-nuts] Linux time servers

[M. Warner Losh]

In message: <20090514220030.68E47BCE6 at ip-64-139-1-69.sjc.megapath.net>
            Hal Murray <hmurray at megapathdsl.net> writes:
: 
: > Is there any consensus for the reasons why Linux performs poorly?  I
: > was thinking about setting up a server as well (possibly using a
: > little ARM-based single-board computer that runs Linux). 
: 
: Consensus?  I doubt it.
: 
: My reading.  Lots of cooks.  None of them are time geeks.

Many of these cooks have firmly held, strong views on how things
should be done, which also gets in the way of having a reasonable
discussion about why those views hurt time performance.  At least
that's been my experience.

On FreeBSD you have phk@ and I which works better...

: There are a lot of people working on Linux.  A lot of them are smart.  A lot 
: of them are not plugged into the culture of key chunks of technology so they 
: "fix" or "clean up" some code in ways that actually breaks things.

Many times the fixes neglect edge cases, or dismiss the need to get
them right at all (like: who cares if the system time is off by a
second, ntpd will steer that out).

: There are "only" a few big screwups that are on my list these days.
:   No PPS support.
:   The TSC calibration code is broken.  (and it's the default mode)
:   The in-kernel NTP support code is broken.
: 
: The last two work, just not quite correctly.  They are close enough so that 
: you probably won't notice any problems unless you are a geek.

True enough...  If you are, you are way too much about it, but if you
don't it doesn't bother you at all...

: One of the things that is driving some of the changes is making
: things work better for low power applications.  The old scheduler
: used to do a bit of work every clock tick (100 HZ to 1000 HZ).  That
: chews up a lot of power if your battery powered system goes to sleep
: when there is nothing to do.  So it seems reasonable to look ahead
: in the scheduler queue and figure out how long until the next time
: there is work to do and sleep until then.

The problem is that these can't easily be turned off...  The other
problem that the tickless stuff starts to expose is that many of these
platforms have counters that can be used for time keeping, but they
wrap too quickly to sleep for long...

Anyway, I'm totally biased on this stuff, so you should take me with a
grain of salt.

Warner