[ntp:questions] Keeping NTP Honest Redux
tglassey at earthlink.net
Mon Nov 9 22:15:24 UTC 2009
> Evandro Menezes <evandro at mailinator.com> writes:
>> On Nov 9, 9:21=A0am, "David J Taylor" <david-tay... at blueyonder.not-this-
>> bit.nor-this.co.uk.invalid> wrote:
>>> Here's the offset:
>>> and here's the rate:
>>> where values for the last two and half days have been overlaid against
>> I think that we can all agree that these charts indicate that the
>> variation is due to temperature variations. Given that, they also
>> indicate that NTP, as is, is too slow to react to temperature
>> variations, taking something like 8 to 10h to compensate for them,
>> even at 64s sampling period, it seems.
I think it more that there was no design for NTP to compensate for local
loading of any kind. For instance if the system is over a set load NTP
should create alarms for the logging to notice when it is more likely
that time-slippage or drift may occur.
As to performance specifications - if you want to create a
temperature/loading type report I would support that but its not part of
the core protocol per se but rather the reference implementation thereof.
>> The reference stratum 1
>> server, Feenix, seems to suffer from the same temperature-induced
>> daily accuracy variations.
>> Unfortunately, Mills will probably deny that this is an issue and will
>> probably refuse revisions to his algorithm. If this is the case,
>> perhaps it's time to add another algorithm alongside Mill's and let to
>> the user whether he prefers Mill's algorithm or an alternative. Which
>> alternative I have no idea, but I'd sure like to decide what's good
>> for me, not Mills.
So you want a seamless application which takes into account all of the
myriad of HW, Memory, and I/O performance constraints? - Thats not NTP
as we know it today and would take a total rewrite of the drafts to
support that use control.
> Well, you do have that choice, only in different names. chrony uses an entirely
> different algorithm (directly estimating the rate and the offset from a sequence
> of measurments).
But it doesnt support Autokey and has no security capabilities so its
pretty much unusable as a source of evidence if there are smart lawyers
> Of course you do not have that choice under Windows ( but
> perhaps someone could port chrony to windows). There are also other places where
> they are not equivalent, but it is at least an option.
The problem with time (MS-SNTP) from the Windows world is how it
controls the selection of the 'most correct' time and how it doesn't
allow for the evidence-trappings of NTP to be promulgated into the A-D
W32TIME.DLL control models.
> questions mailing list
> questions at lists.ntp.org
> No virus found in this incoming message.
> Checked by AVG - www.avg.com
> Version: 8.5.425 / Virus Database: 270.14.57/2492 - Release Date: 11/09/09 12:11:00
More information about the questions