On 10/ 6/16 12:24 PM, Matthew Parsons wrote:
Oh wow, I meant what version of Windows guests! (As you might expect,
guest behavior can be rather different between, say, 2003 vs 2012 R2.)
It's 2012 R2.
Quick google shows that version probably came from
https://download.joyent.com/pub/vmtools/, so hopefully you're runing
2012 :) - It's been a fair few months since I've had to deal w/
Windows under KVM, and it sounds like I don't know anything you don't
- but just in case, I'd start at
https://fedoraproject.org/wiki/Windows_Virtio_Drivers and try updating
w/ the latest there.
I've just migrated the machine back to a single core test host and it is
now using 39% of the host CPU, so that blows the lots of cores
hypothesis. Another 2012R2 guest on the same host is idling at 4%, so
there's something going on the the VM.
The busy qemu process is making a very high number of ioctl calls,
mainly request id 536915584. I'm guessing this is due to the high guest
interrupt rate (10X the other guest.
(http://www.linux-kvm.org/page/WindowsGuestDrivers/Download_Drivers -
for more info)
On Wed, Oct 5, 2016 at 3:48 PM, Ian Collins <[email protected]
<mailto:[email protected]>> wrote:
On 10/ 6/16 10:03 AM, Matthew Parsons wrote:
Have you installed the VirtIO drivers in windows? (And what
version?) For testing I'd try disabling/removing the guest
NICs and just see if interrupts die down.
Which version is often an issue! There are so many out there and
some work with one version of windows and not another! The ISO I
use is named "me-ws2012std-20130712.iso" and I've been using it
for a couple of years so its origins are lost in the mists of time.
Also (again for testing) perhaps reduce cores to the amount on
a physical CPU socket and assign/restrict to avoid crossing
NUMA boundries.
The problem only becomes an issue when the core number gets high,
as I said in my original post the load average almost quadruples
when going from 16 to 32 cores.
(I trust that whatever workload you're running benefits from
that many cores, but typically I'd keep 2 or so for the
hypervisor/management/other.)
The workload is compiling a large C and C++ code base, so the more
cores the better.
Experimenting on a smaller machine shows the build times to cores
ratios reflect those on bare metal, that is if I give the VM the
full system picture (using qemu_extra_opts) build times are about
25% faster than giving it the number of physical cores (using
vcpus). For example to get optimum performance on a single quad
core, use "vcpus": 1, "qemu_extra_opts": "-smp
cpus=1,cores=4,threads=2".
Cheers,
--
Ian.
*smartos-discuss* | Archives
<https://www.listbox.com/member/archive/184463/=now>
<https://www.listbox.com/member/archive/rss/184463/22567596-23bfd832>
| Modify
<https://www.listbox.com/member/?&;>
Your Subscription [Powered by Listbox] <http://www.listbox.com>
--
Ian.
-------------------------------------------
smartos-discuss
Archives: https://www.listbox.com/member/archive/184463/=now
RSS Feed: https://www.listbox.com/member/archive/rss/184463/25769125-55cfbc00
Modify Your Subscription:
https://www.listbox.com/member/?member_id=25769125&id_secret=25769125-7688e9fb
Powered by Listbox: http://www.listbox.com
