konrad at kernel
Aug 31, 2012, 10:42 AM
Post #33 of 34
On Fri, Aug 31, 2012 at 10:36:47AM -0700, George Dunlap wrote:
> On Tue, Aug 28, 2012 at 3:06 AM, Ian Campbell <Ian.Campbell [at] citrix> wrote:
> > * [BUG] qemu-traditional has 50% cpu utilization on an idle
> > Windows system if USB is enabled. Not 100% clear whether this is
> > Xen or qemu. George Dunlap is performing initial
> > investigations.
> So it's hard to get directly comparable results, but I think that
> early indications are that the biggest chunk of this is due to the
> extra syscall overhead for a 64-bit dom0. Data points are:
> 1. Ubuntu 12.04, 64-bit, pvops Ubuntu kernel, Xen 4.2-rc2, older AMD
> system: qemu uses 50% on an idle system
So what happens if you run with a 32-bit dom0? What is the kernel
version? There were some issues with extra traps being done due to the
cpuidle running (which it should not).
> 2. XenServer built with Xen-4.2; (32-bit 2.6.32 dom0), Nehalem system:
> <qemu uses 2% on an idle system
> 3. Debian wheezy with the squeeze 2.6.32 32-bit kernel, older AMD
> system: qemu uses 10% on an idle system
Can you try booting with 'nohz=off'. What does 'perf top' (you need to
run v3.4 or later) give you?
> Looking at the traces, it seems that on the AMD box there were just a
> whole lot more USB-related IO accesses than on the Nehalem system. #2
> had far fewer USB-related accesses than #1, but #3 had about twice as
> many as #1. So it seems likely to be a combination between something
> weird that the USB driver in the guest is doing under AMD, and the
> extra overhead of a 64-bit kernel.
> So I think this is probably OK to take off the blocker list (although
> it's probably something we want to look into further).
> Xen-devel mailing list
> Xen-devel [at] lists
Xen-devel mailing list
Xen-devel [at] lists