"Stephen Gutknecht (SAPDB)" schrieb:
> 
> Hello Joerg,
> 
> You obviously don't work in the same areas I do :)  Our serves must be
> synced to a known references constantly.
> 
> Windows 2000 on intel hardware does not keep that good time.  I notice our
> SMP Intel SCB2 motherboard systems (dual Pentium 3 1.26Ghz) drift fast about
> 30ms every 20 minutes.  The harder the load on the CPU, the more the drift.
> 
> The way we deal with it is to set the clock every 20 minutes ... that way
> the changes are not big enough to foul up log times and so forth.  We wrote
> a custom program that even makes sure the absolute change is never more than
> 1/4 of a second (250ms), that way big changes are avoided and can be done
> during reboot cycles.  Normally the every 20 minute change keeps things in
> sync.

Your problem seems to be a Windows problem. I'm running 23 Linux servers
across Europe connected trough WAN links ranging from 1Mbit/s to as
small as 16kbit/s and they sync time via ntpd against two central
servers. Max. offset is around ~25ms, typical offset is ~10ms. I
remember a project where we had to sync many servers in a similar
environment, and it turned out that the Windows boxes were the difficult
ones. Did you try the Windows version of NTP
http://www.eecis.udel.edu/~ntp/software/winnt.html . I have not used it
but it may be useful.

Simon

> 
> I have seen other systems drift slow... each motherboard/CPU combination has
> a different behavior.
> 
> We have to connect to outside web servers on a 1/2 second (500ms) accuracy
> for one of our business ventures, so it is not an option to let the clock
> drift out.
> 
> Think of it this way:  What if you have 10 servers running the same
> application, logging to a common database, wouldn't it be worse if they all
> had different times?  PC clocks are not the greatest...  Yes, you could let
> the database provide the timestamp, but sometimes you are recording the time
> from an event that happens before the actual SQL INSERT statement.
> 
> I know there are hardware solutions to this (better clock chips and driver
> that references the hardware), but it has just been easier to set the clock
> in software.  Windows and other operating systems I have worked with have
> agents to automate this.
> 
> To answer your questions:
> 
> How do you deal with doubled timed period?  That is where sequence number is
> combined with time period.  In our time-critical application, we can still
> tell what comes first - even if the time stamp shows otherwise (rare).
> 
> I fully understand the hibernation issue - but I really doubt too many
> people are hibernating their servers.  Adjusting the clock is pretty common
> practice that I have seen...  The time zone change you reference, Windows
> practically encourages this.  A  application solution is to use universal
> time, but still... clock corrections do have to be made based on my
> experience.
> 
> Regards,
> 
>   Stephen
> 
> -----Original Message-----
> From: Mensing, Joerg [mailto:[EMAIL PROTECTED]]
> Sent: Wednesday, November 20, 2002 5:24 AM
> To: 'Stephen Gutknecht (SAPDB)'
> Cc: [EMAIL PROTECTED]
> Subject: RE: Problem with hibernate: 7.4 Beta randomly crashing (Windows)
> 
> Hi,
> it concerns my that you think about restting time on a server. You can set
> your clock into the past, while SAPDB is running you, but only using a time
> daemon in differential mode. As SAP customer you are forced to do so or your
> R/3 system may crash.
> 
> SAPDB may survive that, but i would stop it anyhow before i would to reset
> the time!
> 
> Why?
> 
> What about your applications? How do they handle a 'doubled' time period? If
> your application survives daylight saving time changes without logical
> problems (if they use local timestamps), it is save but only very small
> number of applications do really survive daylight saving time without
> logical problems. If you wildly modify your local time you will confuse any
> kind of time depending process. Think about make dependency i.e. or cron
> jobs...
> 
> ... but the hibernation problem is a differrent thing, since it resets the
> internal clock register! On NT you will not influence that internal CPU
> register by setting the wall clock. The register used is only counting the
> external clock chrystal signal depending on the MHz provided, but has
> nothing to do with your wall clock.
> 
> In a good server environment you should try not to play with the time scale.
> 
> CU
> jrg
> 
> > -----Original Message-----
> > From: Stephen Gutknecht (SAPDB) [mailto:[EMAIL PROTECTED]]
> > Sent: Dienstag, 19. November 2002 21:24
> > To: Mensing, Joerg; [EMAIL PROTECTED]
> > Subject: RE: Problem with hibernate: 7.4 Beta randomly crashing
> > (Windows)
> >
> >
> > Humm...
> >
> > This concerns me a bit.  What happens if the clock is changed
> > while SAPDB is
> > running?  Is there any chance that a clock correction (fast
> > clock set back
> > 20 seconds every 12 hours) could cause problems?
> >
> >   Stephen
> >
> >
> > -----Original Message-----
> > From: Mensing, Joerg [mailto:[EMAIL PROTECTED]]
> > Sent: Monday, November 18, 2002 4:52 AM
> > To: '[EMAIL PROTECTED]'; [EMAIL PROTECTED]
> > Cc: Roedling, Raymond
> > Subject: Problem with hibernate: 7.4 Beta randomly crashing (Windows)
> >
> >
> > Hi,
> > with turning hibernation on, Raymond was able to reproduce
> > the crash. It
> > seems to be the hibernate mode then... If hibernation is
> > turned on, the CPU
> > timer register seems to be spontanously resetted during wakeup. This
> > confuses time measurement function, which does not expect to
> > have entered
> > the dispatcher function after leaving it...
> >
> > This does not have an effect both on 'PERFCNTR' and 'TICKS' setting of
> > HIRES_TIMER_TYPE. So if you decide to allow hibernation of your laptop
> > you definitly have to modify the parameter.
> >
> > CU
> > jrg
> >
> >
> > > -----Original Message-----
> > > From: Mensing, Joerg
> > > Sent: Montag, 18. November 2002 09:34
> > > To: '[EMAIL PROTECTED]'; [EMAIL PROTECTED]
> > > Cc: Roedling, Raymond
> > > Subject: RE: 7.4 Beta randomly crashing (Windows)
> > >
> > >
> > > Hi,
> > > in 7.4 a new high resolution timing code is used. The method is
> > > selectable by parameter "HIRES_TIMER_TYPE" (NT only). Default is
> > > 'CPU' which means the code uses the RDTSC instruction to get the
> > > actual tick count from an internal CPU register.
> > > It seems that this code does not work with your laptop CPU.
> > > You should try to set the parameter "HIRES_TIMER_TYPE" to 'PERFCNTR'
> > > or 'TICKS' instead. I hope your random crashes will disappear.
> > > CU
> > > jrg
> > >
> > > > -----Original Message-----
> > > > From: Ray Harrison [mailto:[EMAIL PROTECTED]]
> > > > Sent: Samstag, 16. November 2002 00:20
> > > > To: [EMAIL PROTECTED]
> > > > Subject: 7.4 Beta randomly crashing (Windows)
> > > >
> > > >
> > > > Hello -
> > > > I have downloaded the beta for 7.4 and have it installed on my
> > > > laptop - PIII/500mz 384mb ram, Win2K professional, SP3. I'm not
> > > > really doing anything with it except for the odd table creation,
> > > > etc. Anyway, I have noticed that at random times, I will get
> > > > a message about the kernal aborting
> > > > and shutting down. I have enclosed the error file from
> > > > today's sessions. I also notice the same
> > > > thing on my Proliant 6500 - Win 2K Server, 1.5 gig, but have
> > > > been using the laptop mostly. It may
> > > > that I am not setting everything up correctly of course, but
> > > > I assumed that random crashes
> > > > (seemingly) aren't a good thing!
> > > >
> > > > Enclosed, please find the error file and let me know should I need
> > > > to provide additional information.
> > > >
> > > > Warm Regards -
> > > > Ray Harrison
> > > >
_______________________________________________
sapdb.general mailing list
[EMAIL PROTECTED]
http://listserv.sap.com/mailman/listinfo/sapdb.general

Reply via email to