Xen devel - Mar 2008 - Issues with multiple differing guest HZ rates?

In a timer patch in RHEL5u1 (by Alan Cox), is the comment:

"Xen itself has issues with multiple differing guest GZ rates"
(and I assume that GZ is a typo, should be HZ)

Can anyone explain this further or are there experimental
results to back it up?  We are seeing some strangeness with
running multiple pvm guests with different HZ rates and wonder
if this is related.

Thanks,
Dan


==================================If Xen could save time in a bottle / then
clocks wouldn't virtually skew /
It would save every tick / for VMs that aren't quick /
and Xen then would send them anew
(with apologies to the late great Jim Croce)
_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Dan Magenheimer wrote:
> In a timer patch in RHEL5u1 (by Alan Cox), is the comment:
> 
> "Xen itself has issues with multiple differing guest GZ rates"
> (and I assume that GZ is a typo, should be HZ)
> 
> Can anyone explain this further or are there experimental
> results to back it up?  We are seeing some strangeness with
> running multiple pvm guests with different HZ rates and wonder
> if this is related.
I highly doubt it; in that comment, Alan was describing why he *didn''t*
implement the tick divider for Xen kernels.  So that patch doesn''t
actually touch any Xen code in particular, and should in general not be
touching much code that the -xen kernels share with the bare-metal
kernel either.  What''s the problem you are seeing with different HZ
rates in guests?

Chris Lalancette

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel

Hi Chris --

We''ve noticed that a 100Hz RHEL5-64bit kernel (compiled 100 HZ, not
divider=10) has less clock skew with timer_mode=0 than with
timer_mode=2.  This would be the only such we''ve found for
a 64-bit kernel and this is worrisome because it complicates
the selection algorithm for setting timer_mode.  In our clock skew
tests, we run six 2vcpu 2GB guests on a 8pcpu 16GB system; this
was the only guest with 100Hz tick, all the rest are 1000Hz.
timer_mode=2 works fine if that guest is 1000Hz.

We''ll continue to investigate but Alan''s comment in that patch
was intriguing enough to ask for more info.

Thanks,
Dan
> -----Original Message-----
> From: Chris Lalancette [mailto:clalance@redhat.com]
> Sent: Sunday, March 02, 2008 8:14 PM
> To: dan.magenheimer@oracle.com
> Cc: xen-devel@lists.xensource.com
> Subject: Re: [Xen-devel] Issues with multiple differing guest 
> HZ rates?
> 
> 
> Dan Magenheimer wrote:
> > In a timer patch in RHEL5u1 (by Alan Cox), is the comment:
> >
> > "Xen itself has issues with multiple differing guest GZ
rates"
> > (and I assume that GZ is a typo, should be HZ)
> >
> > Can anyone explain this further or are there experimental
> > results to back it up?  We are seeing some strangeness with
> > running multiple pvm guests with different HZ rates and wonder
> > if this is related.
> 
> I highly doubt it; in that comment, Alan was describing why 
> he *didn''t*
> implement the tick divider for Xen kernels.  So that patch doesn''t
> actually touch any Xen code in particular, and should in 
> general not be
> touching much code that the -xen kernels share with the bare-metal
> kernel either.  What''s the problem you are seeing with different
HZ
> rates in guests?
> 
> Chris Lalancette
> 

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xensource.com
http://lists.xensource.com/xen-devel