Notes on Windows 10 Time Service

There are some very useful changes to the available registry entries to control Windows Time Service in recent W10 versions:

  • Slewing during leap-second events
  • Maximum slew rates for corrections
  • Spike detection thresholds
  • Max/min polling intervals
  • Configurable update interval during slews
  • Verbose logging.

I have configured it pretty much the same as I had configured the NTP from Meinberg, with a set of UK-based pool.ntp.org servers, and a minimum poll of 64s and max poll of 1024s.

At initialisation, it does a DNS lookup for the hostnames of configured time servers, returning multiple addresses as expected when I use the 0.uk.pool.ntp.org address (or the 1., 2. or 3. of course).

It then sends a single NTP request to the first address in each of the pools, using NTP version 3.

and gets back a response:

It then polls again after 64 seconds.  Once the dispersion and offset are within limits, it backs off to 128s, then 256s and finally to 1024s between polls.  Only a single request each time.

Initially, the HostPoll value is 64s.

After 20 minutes, it has backed off to 256s between polls

Checking sync with other ntp servers out in the internet, it looks like we are within 10ms, but again that is probably closer than you could verify using a simple request as the network latency is at least twice that.

The slewing controls need some serious verification, as does the spike detection.  For an SNTP server, this isn’t too bad.  It does look as though there isn’t quite the equivalent of the drift file in NTP, but there is a record in the registry showing the current clock rate, which can be offset from the default 156250 and may offer the same drift-control facility (mine is at 156249 at the moment).

In HKLM\SYSTEM\CurrentControlSet\Services\W32Time\Config I have:

I have enabled verbose logging, checked the HoldPeriod is set to the default 5 attempts and set the Max and Min poll intervals to 2^10 (1024s) and 2^6 (64s) respectively.

The UpdateInterval only applies where you have configured a Type 1 NTP server, and there is a special update interval setting to match that.  I am using only Type 3 server entries (the default) which says “Dear Server, please feel free to configure my time, but I won’t try to mess with yours”.

In the HKLM\SYSTEM\CurrentControlSet\Services\W32Time\TimeProviders\NtpClient registry section I have:

SpecialPollInterval isn’t used except with a type 1 server entry.

The key

Blocking the replies from the currently-selected source results in the ValidDataCounter for that source decrementing by one after each update interval.  Despite the update interval being dropped to 16s, the W32Time client process only appears to poll at the rate of the overall application, and it is now 256s, so I have a long wait for it to time out.  It has decremented the reachability to zero and the same for the ValidDataCounter and puit the server intp “Pending” status after about 1000s.  No sign of any DNS lookups since the client first started.  Also (weirdly) no sign of a DNS TTL in the DNS returned by my broadband router, and neither is there any entry in ipconfig /displaydns for the time servers.

After 2400 seconds, the client decided that the server was dead, and it did another DNS lookup for the missing host, and started sending single ntp queries to the new IP address.

I’ll keep an eye on it to see if it remains at this sort of offset. 10 to 20ms is fine for all the hobby digimode stuff I’m doing, but it would be pants for forensics or trading. After a few hours is it running at 5ms ahead of a range of other ntp servers, climbing slowly.  It hit 8ms and the clock value changed and now it is ramping slowly downwards.

So at the moment, this up to date Win10 1903 desktop is running a slightly-more-configurable SNTP with decent audit logging, but not “proper” NTP.  Is there a trick to make it do “proper” NTP like Meinberg does?  I tried making the servers type 0x1, that had no effect.

Also, when a normal human being is asked to make a load of registry changes just to get a useable time facility, the response is usually a bit less positive than “Oooh, great, no problems”.

Can’t say I’m over-keen on putting the necessary keys into a .reg file for distribution, but I guess that is the only way except for reg-hacking geeks.

Meinberg beats this hands-down for simplicity and performance at the moment, and the Meinberg monitoring tool and CLI tools are better. OK, I could run a load of CLI things in Powershell or the Windows Linux subsystem, but it still feels like the usual MS way of reinventing the wheel, late and poorly.

We shall see if the stability and handling of spiked or drifty sources is better or worse than “proper” NTP. 

Leave a Reply

Your email address will not be published.