Poor write- and overall performance

Poor write- and overall performance

[Arne Kepp]

Hi,

I'm testing KVM 80 (prepackaged from lfarkas.org) on CentOS 5.2 (both 
guest and host) and comparing against Xen 3.3.0. The only modification 
I've really made is that I've set noatime on both guest and host. The 
guest is running straight from an LVM volume and is assigned 4 Gb RAM 
and 4 virtual CPUs.

KVM does well reading from disk (300 Mbyte/s according to IOZone, 
virtually the same as native) and computational tasks like OpenSSL 
signing.  But writing ends up at 30 Mbyte/s (native does 150 Mbyte/s) 
and kernel compilation times are roughly double that of native. KVM 79 
exhibited the same behavior for disk IO, but froze while compiling the 
kernel.

So, short of using the virtio drivers (which I think would require a 
custom kernel since RHEL 5.2 uses 2.6.18), are there any easy tweaks I 
should try to make KVM fare better ?

The host is quad core Intel with 8 Gb RAM and a SAS RAID 5 
configuration. Let me know if you'd like to see the results here.

Sincerely,
-Arne

-- 
Arne Kepp
OpenGeo - http://opengeo.org
Expert service straight from the developers

Re: Poor write- and overall performance

[Avi Kivity]

Arne Kepp wrote:
> Hi,
>
> I'm testing KVM 80 (prepackaged from lfarkas.org) on CentOS 5.2 (both 
> guest and host) and comparing against Xen 3.3.0. The only modification 
> I've really made is that I've set noatime on both guest and host. The 
> guest is running straight from an LVM volume and is assigned 4 Gb RAM 
> and 4 virtual CPUs.

I suggest trying scsi in kvm-81, or setting cache=writeback if you want 
to keep using IDE (but be aware of the potential for severe data loss 
with the latter option).

-- 
I have a truly marvellous patch that fixes the bug which this
signature is too narrow to contain.

Re: Poor write- and overall performance

[Henrik Holst]

mån 2008-12-15 klockan 08:38 +0200 skrev Avi Kivity:
> Arne Kepp wrote:
> > Hi,
> >
> > I'm testing KVM 80 (prepackaged from lfarkas.org) on CentOS 5.2 (both 
> > guest and host) and comparing against Xen 3.3.0. The only modification 
> > I've really made is that I've set noatime on both guest and host. The 
> > guest is running straight from an LVM volume and is assigned 4 Gb RAM 
> > and 4 virtual CPUs.
> 
> I suggest trying scsi in kvm-81, or setting cache=writeback if you want 
> to keep using IDE (but be aware of the potential for severe data loss 
> with the latter option).
> 
Now I haven't tried kvm-81 but scsi with 4GB of RAM or more is severely
broken in QEMU for all other tested KVM versions (80 and downwards).

/Henrik Holst

Re: Poor write- and overall performance

[Avi Kivity]

Henrik Holst wrote:
> Now I haven't tried kvm-81 but scsi with 4GB of RAM or more is severely
> broken in QEMU for all other tested KVM versions (80 and downwards).
>   

kvm-81 contains this:


>     Add 40-bit DMA support to LSI scsi emulation (Ryan Harper)
>     
>     This patch fixes Linux machines configured with > 4G of ram and 
> using a
>     SCSI device.

Which should fix the problem.

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

Re: Poor write- and overall performance

[Henrik Holst]

mån 2008-12-15 klockan 12:26 +0200 skrev Avi Kivity:
> Henrik Holst wrote:
> > Now I haven't tried kvm-81 but scsi with 4GB of RAM or more is severely
> > broken in QEMU for all other tested KVM versions (80 and downwards).
> >   
> 
> kvm-81 contains this:
> 
> 
> >     Add 40-bit DMA support to LSI scsi emulation (Ryan Harper)
> >     
> >     This patch fixes Linux machines configured with > 4G of ram and 
> > using a
> >     SCSI device.
> 
> Which should fix the problem.
> 
Now this is good news indeed!

/Henrik Holst

Re: Poor write- and overall performance

[Arne Kepp]

Avi Kivity wrote:
> Arne Kepp wrote:
>> Hi,
>>
>> I'm testing KVM 80 (prepackaged from lfarkas.org) on CentOS 5.2 (both 
>> guest and host) and comparing against Xen 3.3.0. The only 
>> modification I've really made is that I've set noatime on both guest 
>> and host. The guest is running straight from an LVM volume and is 
>> assigned 4 Gb RAM and 4 virtual CPUs.
>
> I suggest trying scsi in kvm-81, or setting cache=writeback if you 
> want to keep using IDE (but be aware of the potential for severe data 
> loss with the latter option).
>

Thanks for the tip. I tested with the SCSI driver and 81, unfortunately 
write performance was approximately the same as before.

Keep up the good work,
-Arne

-- 
Arne Kepp
OpenGeo - http://opengeo.org
Expert service straight from the developers

Re: Poor write- and overall performance

[Ryan Harper]

* Arne Kepp <ak@opengeo.org> [2008-12-19 19:45]:
> Avi Kivity wrote:
> >Arne Kepp wrote:
> >>Hi,
> >>
> >>I'm testing KVM 80 (prepackaged from lfarkas.org) on CentOS 5.2 (both 
> >>guest and host) and comparing against Xen 3.3.0. The only 
> >>modification I've really made is that I've set noatime on both guest 
> >>and host. The guest is running straight from an LVM volume and is 
> >>assigned 4 Gb RAM and 4 virtual CPUs.
> >
> >I suggest trying scsi in kvm-81, or setting cache=writeback if you 
> >want to keep using IDE (but be aware of the potential for severe data 
> >loss with the latter option).
> >
> 
> Thanks for the tip. I tested with the SCSI driver and 81, unfortunately 
> write performance was approximately the same as before.

Yeah, currently the scsi emulation doesn't queue up writes, rather it
waits for each one to complete which means one cannot take advantage of
the aio infrastructure.  I'm currently working on fixing this; stay tuned.

-- 
Ryan Harper
Software Engineer; Linux Technology Center
IBM Corp., Austin, Tx
ryanh@us.ibm.com