[asterisk-users] hardware clock drift and CDR

[Gordon Henderson]

On Mon, 26 Apr 2010, Vieri wrote:

> --- On Sun, 4/25/10, Gordon Henderson <gordon+asterisk at drogon.net> wrote:
>
>>> Hi,
>>>
>>> I've noticed that one of my new servers (new mobo) if
>> drifting slowly
>>> backwards in time (in aprox. 24 hours, system time
>> drifts back 5
>>> minutes).
>>>
>>> I have an ntpd process which is supposed to sync with
>> a lan time server
>>> but it's not quite working. So I'm launching a manual
>> ntpdate or
>>> ntp-client once an hour and that seems to work.
>>
>> If you can run ntpdate and it sets the time, then you are
>> not running
>> ntpd. The 2 can not run at the same time.
>
> Hi Gordon,
>
> Are you sure about this?

Yes.

>ntpd is a daemon and adjusts the time in a continuous manner. ntp-client 
>or ntpdate or whatever are one-time clients that reset the system clock. 
>I don't see why an ntp-client can't be run while ntpd is working (it
>shouldn't be necessary but may come in handy when the time difference is 
>big and ntpd refuses to sync).

ntp binds to the ntp port (123) and prevents anything else binding to it, 
or listening on it - which ntpdate needs to do.

Example here:

Desktop is running ntpd:

   yakko:/home/gordon# ps ax | fgrep ntp
   22064 ?        Ss     0:14 /usr/sbin/ntpd -p /var/run/ntpd.pid -u 106:107 -g
   30340 pts/29   R+     0:00 fgrep ntp

I try to run ntpdate:

   yakko:/home/gordon# ntpdate essen.drogon.net
   26 Apr 14:20:47 ntpdate[30341]: the NTP socket is in use, exiting

> Anyway, I've noticed that my ntpd log messages don't say "anything" when 
> trying to sync to my "Windows PDC LAN time server". Curiously, 
> ntp-client DOES sync to this Windows server.

> So I decided to sync to pool.ntp.org and now I see syslog messages that 
> actually show that the system time gets adjusted by ntpd.
>
> I'd rather sync to my LAN time server but this is off-topic on this ML.

Using pool and your LAN server would be the best way forward - there are 
pool server avalable for most countries too, so us.pool.ntp.org, 
uk.pool.ntp.org, and so on.

Your /etc/ntp.conf file can be very simple indeed - my workstation one is 
nothing more than:

   server essen.drogon.net
   server  uk.pool.ntp.org

You can check your servers ntp daemon with:

   ntpq -c peers

and

   ntpq -c rl

The key thing to look for in the 'rl' command is 'stratum'. If it's 16 
then it's not synchronised and anything less than 16 is good.

   yakko:/home/gordon# ntpq -c rl | fgrep stratum
   processor="i686", system="Linux/2.6.29.2", leap=00, stratum=4,

Don't get too hung-up on how close to zero the stratum is.

>>> How does Asterisk CDR count the duration/billsec
>> values? Does it rely on
>>> system time ONLY for "call start" or also for "call
>> end"?
>>>
>>> What Asterisk-related side-effects should I expect
>> from a drifting
>>> clock?
>>
>> Who cares. Just fix ntpd then your worys are gone.
>
> Well, I still have doubts about that. I could look at * source code but 
> I'd rather hear from someone here.

Might be easier to read the code ;-)

> My ntp log shows this:
>
> 26 Apr 13:06:30 ntpd[534]: synchronized to xxx.xxx.xxx.xxx, stratum 2
> 26 Apr 13:21:24 ntpd[534]: time reset +2.318647 s
> 26 Apr 13:21:44 ntpd[534]: synchronized to xxx.xxx.xxx.xxx, stratum 2
> 26 Apr 13:37:46 ntpd[534]: time reset +2.325417 s
> 26 Apr 13:38:06 ntpd[534]: synchronized to xxx.xxx.xxx.xxx, stratum 2
> 26 Apr 13:54:11 ntpd[534]: time reset +2.327974 s
> 26 Apr 13:55:19 ntpd[534]: synchronized to xxx.xxx.xxx.xxx, stratum 2
> 26 Apr 14:09:16 ntpd[534]: time reset +2.177572 s
> 26 Apr 14:10:08 ntpd[534]: synchronized to xxx.xxx.xxx.xxx, stratum 2
> 26 Apr 14:26:07 ntpd[534]: time reset +2.357017 s
>
> That kind of scares me because if I'm not mistaken it means that about 
> every 20 seconds, my ntpd adjusts the system time by about 2 seconds 
> forward. So my clock is going back 2 seconds every 20... That's a 
> significant drift. And it would definitely make a difference in my CDR 
> records IF Asterisk were to compare the "start and end" system times.
>
> Should I worry about this?

If ntpd can't keep the kernel time in-sync then it will step abput every 
900 seconds - which is what appears to be happening here. (the intervals 
are typically much longer than 20 seconds - e.g. 13:06:30 to 12:21:24 is 
~15 minutes - 900 seconds.

I don't think I've ever had a server a bad as that before, so have never 
looked further... Still, it's 2 seconds in 900 seconds, not 2 in 20 as you 
thought.

Which I think is odd - the Linux clock is software derived based on a 
hardware interrupt - it only consults the hardware battery-backed clock at 
boot time (and is supposed to write the current time to it at shutdown 
time) so I wonder if your server is missing interrupts, or otherwise 
mis-behaving.

Is there anything else odd in the log-files?

Gordon