From: Avi Kivity <avi@redhat.com>
Subject: Re: Poor KVM guest performance on an HP rack server
Date: Sun, 13 Dec 2009 12:13:38 +0200
Message-ID: <4B24BE52.6060703@redhat.com>
References: <4B243B2B.806@pardus.org.tr>
Mime-Version: 1.0
Content-Type: text/plain; charset=UTF-8;
	format=flowed
Content-Transfer-Encoding: QUOTED-PRINTABLE
Cc: "linux-kernel@vger.kernel.org" <linux-kernel@vger.kernel.org>,
	KVM list <kvm@vger.kernel.org>
To: =?UTF-8?B?T3phbiDDh2HEn2xheWFu?= <ozan@pardus.org.tr>
In-Reply-To: <4B243B2B.806@pardus.org.tr>
Sender: kvm-owner@vger.kernel.org

On 12/13/2009 02:54 AM, Ozan =C3=87a=C4=9Flayan wrote:
> Hi,
>
> We have an HP Proliant DL580G5 rack server. It has 4 Intel Xeon X7460=
(6
> core, 2.67GHz, 16MB L3) processor with 32GB of memory. /proc/cpuinfo =
has
> 24 x the following entry:
>
> I'm running 2.6.30.9-pae on top of it. We were actually planning to u=
se
> it for a virtualization server for giving people dedicated *guest*
> access for their personal compile-farm needs.
>   =20

=46irst, as Jeremy notes, large memory machines want an x86_64 kernel. =
=20
But that doesn't explain the slowness.

> For testing purposes, we created a guest VM (2.6.30.9 too) on top of =
it
> with 2GB of virtual memory stored in a raw partition:
>
> qemu-kvm -cpu host -smp 2 -m 2047 -drive
> file=3D/dev/cciss/c1d0p1,if=3Dvirtio,cache=3Dnone,boot=3Don -net
> nic,model=3Dvirtio,macaddr=3DDE:AD:BE:EF:10:28 -net
> tap,ifname=3Dtap0,script=3D/usr/bin/qemu-ifup -k tr -nographic -daemo=
nize
>
>
> The problem is that I'm seeing very poor performance within the guest=
=2E
> I've googled a bit and seen similar bug reports/discussions ending wi=
th
> some tweaks (setting rotational to 1 for virtio_blk, using cache=3Dno=
ne,
> etc.) and an explanation from Avi Kivity about the bad scalability of
> KVM on pre-Nehalem boxes under high build load.
>   =20

-smp 2 should work perfectly well.

> But I fear that I'm far behind that *bad scalability*. I've made some
> comparisons with my QuadCore Q8300 (2MB cache) box. I won't give the
> whole numbers but for example,
>
> Running the autotools configure script of CUPS on that KVM guest (I c=
an
> easily follow the output of configure line per line, it really really
> waits on some checks):
>
> real    0m52.876s
> user    0m4.892s
> sys     0m55.705s
>
> On the host (while running the guest vm):
>
> real    0m8.193s
> user    0m3.099s
> sys     0m4.055s
>
> On the quadcore box:
>
> real    0m8.424s
> user    0m2.651s
> sys     0m2.879s
>
> Both with cold cache (echo 3>  /proc/sys/vm/drop_caches)
>
> So it's not even a high build load. I've tried with -smp 8 (which sho=
wed
> worse numbers than -smp 2 and 4), with IDE instead of virtio, without
> -cpu host parameter but can't get near 30 (I've got 35 seconds with
> tuning read_ahead_kb, on top of IDE instead of virtio, etc.) seconds =
at all.
>
> I've also tried hugetlbfs for backing the memory within the guest.
>
> I'm using the latest kvm-mod-2.6.32 built on top of 2.6.30.9.
>
> So is this huge performance difference should be accepted as normal o=
r
> am I missing some big things?
>   =20

=46irst, are you sure that kvm is enabled?  'info kvm' in the monitor.

Second, is the workload cpu bound or I/O bound?  Both from the guest's=20
and host's point of view.

--=20
error compiling committee.c: too many arguments to function