[ntp:questions] Red Hat vote for chrony
cswiger at mac.com
Tue Dec 9 07:06:19 UTC 2014
On Dec 8, 2014, at 12:04 PM, Phil W Lee <phil at lee-family.me.uk> wrote:
>> With air-conditioning on, the temperature change shrinks to about 5C,
>> which reduces the thermal wandering of the XO by a factor of 2. That
>> seems to be a worthwhile improvement, not "largely irrelevant".
> Of course, server class machines tend to be designed with far tighter
> control of temperature, since wide temperature fluctuations are the
> worst enemy of reliability.
Indeed, although some servers are designed for SOHO and home media
stuff, and folks want those to run a little quieter than the typical
datacenter server, which tends to resemble a wind tunnel. :-)
[ ... ]
>> Furthermore, most of the systems I deal with at work are either 1U or
>> blades in datacenter racks with the raised floor forming a plenum to
>> deliver temperature-controlled air to each rack, hot and cold isle design, etc.
>> I can't observe even a 1C change in temp just by running an individual machine
>> or VM at peak load. Even firing off something which causes a load spike for
>> an hour or two across all of the systems or VMs in a particular rack only
>> causes a 2-3C change.
>> In practice, that limits thermal wandering due to load from 10-20 PPM to
>> around ~2 PPM.
> So an order of magnitude better, even when tested to extremes which
> are well beyond any reasonable expectation of "normal use".
> I would expect normal use probably stays within 1 PPM on those
> systems, based on the figures you've given for extreme load
Yes, exactly. Without doing anything unusual, a machine racked
in a DC is likely to have quite reasonable thermal stability.
That said, I'll admit that ~1-2 PPM isn't especially accurate; it'd
be an error of at least tens of milliseconds per day if the local clock
lost connection to a local or networked timesource. That would
translate into weeks or even a month or so before the error accumulated
to a full second, which isn't too bad from a human timescale perspective.
>>> One of my collegues devised a script whose sole purpose was to stress
>>> the cpus so he could use the air coming out to dry his socks.
>>> Ie, cpu load drives temp change which produces time shifts.
>> Yes. That matters the most for freestanding machines which are not kept
>> in air-conditioning. It matters very little for machines in a data center,
>> because the ambient thermals there are controlled fairly precisely.
> And the machines themselves are better designed to run at constant
> temperature irrespective of load.
> Freestanding office machines have noise reduction well above
> temperature stability on the list of design priorities.
> The attitude with them is just to stop it actually overheating to the
> point of failure, although it may struggle with even that when a bit
> of ordinary office dust gets into the works.
> Of course, a server room is unlikely to have very much of that dust,
Yes, I also find it a bit surprising than modern desktop CPUs and GPUs
are willing to run right up to their thermal trip points of ~80 C or so
rather than bump up fan speed a little more to keep them more around 50 C.
Older systems tended to use more aggressive cooling, especially laptops.
Well, smarter firmware and Hall effect sensors to measure fan speed means
you can spin the fans more slowly than if you needed to apply 40% minimum
speed just to be sure that the fan would spin up from idle.
[ ... ]
> Of course, in an ideal system designed for clock stability, the clock
> variation with temperature would be measured, and an offset based on
> the temperature of the crystal applied directly to the output of the
> clock itself, before it was supplied to the operating system.
Yes; you're describing calibrating a temperature-compensated XO, or TCXO.
> In theory, this wouldn't be expensive if done at the mass production
> stage, but clock stability isn't high enough on the design priorities
> for designers to put it into mass market machines.
Nope. Generic AT crystals are dirt-cheap so I doubt many folks would
pay even $10 extra for a TCXO variant, much less be willing to pay
twice as much for their motherboard to get an SC cut crystal in an OCXO.
More information about the questions