From mboxrd@z Thu Jan 1 00:00:00 1970 From: Dave Winchell Subject: Re: (progress on hpet accuracy) and Re: [PATCH] Add a timer mode that disables pending missed ticks Date: Wed, 05 Mar 2008 10:06:55 -0500 Message-ID: <47CEB70F.9010101@virtualiron.com> References: Mime-Version: 1.0 Content-Type: text/plain; charset=ISO-8859-1; format=flowed Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: List-Unsubscribe: , List-Post: List-Help: List-Subscribe: , Sender: xen-devel-bounces@lists.xensource.com Errors-To: xen-devel-bounces@lists.xensource.com To: Keir Fraser , "dan.magenheimer@oracle.com" Cc: Dave Winchell , "xen-devel@lists.xensource.com" , Deepak Patel List-Id: xen-devel@lists.xenproject.org Keir, Dan: This is an update on the hpet work. With limited testing, the accuracy of this method appears to be at least 10 times better than the pit/tsc method for usex loads. Also, as mentioned before, it does not have the time going backwards problem. Another interesting property is that the same policy is used for a 32 bit Linux guest and 64 bit Linux. Two recent 14-16 hour tests were run, one with usex e48 and the other with usex b48 as loads. The same load was run on three 8 vcpu guests on an 8 cpu platform. No ntpd running. ntpdate -b used for initial setting of clock and ntpdate -q used for monitoring drift. The guests were 4u464, 4u564, 4u432 Linux. The results are: test duration drift (secs) 4u464,4u564,4u432 drift % usex b48 16 hrs -.68, -.60, -.68 -.0012 usex e48 15 hrs -.58, -.55, -.58 -.0011 The drift with pit/tsc as checked in reported a few months ago: >> The error tests confirm the equivalence. With overnight cpu loads, >> the checked in version was accurate to +.048% for sles >> and +.038% for red hat. Recall that .05% is the goal for accuracy so that ntpd can synchronize. With pit/tsc we are rather close to that limit. So far, with hpet, we are well under that goal. As for the code structure, I tried layering on vpt.c and had trouble doing so. Therefore, I focused on a direct approach. I'll describe the approach shortly, after I run tests with other loads and guests. Regards, Dave Keir Fraser wrote: >On 26/2/08 14:45, "Dave Winchell" wrote: > > > >>>If this is an integration of hpet into the vpt.c infrastructure then that >>>would be very welcome. >>> >>> >>> >>So far it is not, however I may head in that direction. >>Last night's test had an error of .065% on one guest so I still have some >>work to do. >> >> > >Then I hope that the accuracy improvements have come from a set of changes >whose effects are both explicable and generally for the good (rather than >artefacts of how a particular sub-point release of Linux drives the HPET). > > -- Keir > > > >