Re: CPU time off by a factor of two
On Thu, 24 Nov 2005 18:42:40 +0800, Uwe Dippel wrote: Another problem on the Proliant ML350: the time isn't calculated properly. I noticed the problem with (open)ntpd; which continuously wanted to adjust by thousands and tens of thousands of seconds; after only a few hours of running. So I turned it off and rebooted. Still, the time runs around twice as fast as it should. For completeness, I really want to add that the problem is solved when using 4.0. Thanks ! Uwe
Re: CPU time off by a factor of two
On Mon, 28 Nov 2005 22:15:23 +0100, scorch wrote: intel powerstep or any other garbage -- try turning it off in the BIOS? Do me a favour and tell me where ! - I went through all the many possible settings, but none was any close to this. Except of 'Power Regulator to Proliant', eventually. And this one is 'Disabled', of course. Since it has to make with the number of CPUs, here are the four settings: Processor Hyper Threading HW Prefetcher Adjacent Sector prefetch No-Execute Memory Protection Question to the experts: Could these make a serious difference ? If yes, which should I change ? Uwe
Re: CPU time off by a factor of two
intel powerstep or any other garbage -- try turning it off in the BIOS? cheers, scorch -- out of the frying pan and into the fire
Re: CPU time off by a factor of two
On Fri, 25 Nov 2005 15:43:52 +0800, Uwe Dippel wrote: It's a bug, so it seems now. Sorry, last night I didn't have access so my answer is late: I simply rebooted to single-CPU-kernel; compiled by myself, just as well, and it runs like hell. Exact, I mean. Not a single second off after three hours. (so says ntpdate.) I rebooted again, .mp, and it ran half as fast as it was supposed to. Off-list I was asked to try the one compiled by [EMAIL PROTECTED] I did, good idea, thanks for the suggestion. But the behaviour is identical: half-speed of time calculation. Now I'm back on single, for the time being, that's better than the screwed time. Uwe
Re: CPU time off by a factor of two
...on Fri, Nov 25, 2005 at 03:43:52PM +0800, Uwe Dippel wrote: It's a bug, so it seems now. Sorry, last night I didn't have access so my answer is late: I simply rebooted to single-CPU-kernel; compiled by myself, just as well, and it runs like hell. Exact, I mean. Not a single second off after three hours. (so says ntpdate.) That sounds quite a bit like what I remember reading about, something like the TSC might run at different speeds on different cores depending on thermal throttling, SpeedStep, ACPI state, whatever. So if you're switching to a counter on another core without taking that into account, you're in trouble. Which doesn't mean that's the problem here :) Alex. (no, I actually have no idea what I'm talking about)
Re: CPU time off by a factor of two
On Thu, 24 Nov 2005 14:40:59 -0600, J Moore wrote: I'm not clear on something... does the time drift with OpenBSD *alone*; i.e. without ntpd running? Yes. Now I let it run on its own, without ntpd, through the last night and send me a mail (I am not at it) each minute. Starting at 19.00 exactly, with a single rdate -s, at 23.00 the mail said 21.00. So the factor is 2, pretty exact. When I get to the box later, I will reboot with single CPU kernel. Since it goes into production today, I need it to be up. If necessary, I will have to rdate it once per minute to advance its clock by 30 seconds each minute. I consider this quite a bug, by now. Uwe