[Openstack] Very poor windows 2008 R2 network performance (iperf)

Adam Lawson alawson at aqorn.com
Fri May 9 18:42:11 UTC 2014


Factors: enabling TCP chimney and RSS.


*Adam Lawson*
AQORN, Inc.
427 North Tatnall Street
Ste. 58461
Wilmington, Delaware 19801-2230
Toll-free: (844) 4-AQORN-NOW
Direct: +1 (302) 268-6914



On Fri, May 9, 2014 at 11:41 AM, Adam Lawson <alawson at aqorn.com> wrote:

> Try this Cisco white paper. *10Ge Connectivity with Windows Servers*
>
>
> http://www.cisco.com/c/dam/en/us/products/collateral/switches/nexus-7000-series-switches/C07-572828-00_10Gb_Conn_Win_DG.pdf
>
> See page 14.
>
>
> *Adam Lawson*
> AQORN, Inc.
> 427 North Tatnall Street
> Ste. 58461
> Wilmington, Delaware 19801-2230
> Toll-free: (844) 4-AQORN-NOW
> Direct: +1 (302) 268-6914
>
>
>
> On Fri, May 9, 2014 at 11:09 AM, JR <botemout at gmail.com> wrote:
>
>> Darn. That dropped the performance by ~50% ...
>>
>>
>> On 5/9/2014 1:57 PM, Adam Lawson wrote:
>> > I just heard back. Within Windows, disable Large File Offload.
>> >
>> > Let me know how that goes?
>> >
>> > Mahalo,
>> > Adam
>> >
>> >
>> > *Adam Lawson*
>> > AQORN, Inc.
>> > 427 North Tatnall Street
>> > Ste. 58461
>> > Wilmington, Delaware 19801-2230
>> > Toll-free: (844) 4-AQORN-NOW
>> > Direct: +1 (302) 268-6914
>> >
>> >
>> >
>> > On Fri, May 9, 2014 at 10:54 AM, JR <botemout at gmail.com> wrote:
>> >
>> >> Adam,
>> >>
>> >> If I'm looking in the right place (the redhat virtio ethernet adapter
>> >> properties), there is nowhere to specify forcing the speed or duplex.
>> >> The only value I see is Init.ConnectionRate (which is 10G). I've been
>> >> able to see performance in excess of 2Gb/sec (when running an iperf
>> >> against the ubuntu host on which the VM runs) so it doesn't think it's
>> a
>> >> 1G NIC; the performance is just very poor. The same iperf from a centos
>> >> VM to its host gives > 9Gb/sec.
>> >>
>> >> Thanks
>> >> JR
>> >>
>> >> On 5/9/2014 1:19 PM, Adam Lawson wrote:
>> >>> Is the duplex setting and speed set on each side to force 10ge (since
>> >>> auto-neg seems to not work in this scenario)? Still waiting to hear
>> back
>> >> on
>> >>> steps taken in the situation I described earlier.
>> >>>
>> >>> Mahalo,
>> >>> Adam
>> >>>
>> >>>
>> >>> *Adam Lawson*
>> >>> AQORN, Inc.
>> >>> 427 North Tatnall Street
>> >>> Ste. 58461
>> >>> Wilmington, Delaware 19801-2230
>> >>> Toll-free: (844) 4-AQORN-NOW
>> >>> Direct: +1 (302) 268-6914
>> >>>
>> >>>
>> >>>
>> >>> On Fri, May 9, 2014 at 9:25 AM, JR <botemout at gmail.com> wrote:
>> >>>
>> >>>> I'd be very appreciative to hear how you solved this Adam.  ;-)
>> >>>>
>> >>>> On 5/9/2014 12:29 AM, Adam Lawson wrote:
>> >>>>> Look at the TCP stack within Windows and optimizations recommended
>> by
>> >>>>> Microsoft. I don't think it's a KVM or openstack question to be
>> honest.
>> >>>> We
>> >>>>> ran into similar issues on plain old Win2008 R2 servers that were
>> >> running
>> >>>>> on bare metal. Will update again when I ping someone to find out
>> what
>> >>>>> specifically it was back then.
>> >>>>>
>> >>>>>
>> >>>>> *Adam Lawson*
>> >>>>> AQORN, Inc.
>> >>>>> 427 North Tatnall Street
>> >>>>> Ste. 58461
>> >>>>> Wilmington, Delaware 19801-2230
>> >>>>> Toll-free: (844) 4-AQORN-NOW
>> >>>>> Direct: +1 (302) 268-6914
>> >>>>>
>> >>>>>
>> >>>>>
>> >>>>> On Thu, May 8, 2014 at 6:59 PM, JR <botemout at gmail.com> wrote:
>> >>>>>
>> >>>>>> Greetings,
>> >>>>>>
>> >>>>>> My openstack grizzly cluster runs on ubuntu 12.04 servers with 10G
>> >> NICs.
>> >>>>>> I have ubuntu, centos and windows 2008 R2 guests.  I've noticed
>> that
>> >>>>>> while both my linux guests can communicate at some reasonable
>> >>>>>> approximation of 10G wire speed (e.g., 7-9Gb/sec on iperf tests),
>> the
>> >>>>>> windows guests max out at 2.5G when talking to the host on which
>> they
>> >>>>>> run, down to ~1.2G to other hosts.
>> >>>>>>
>> >>>>>> I've made some modifications as per this doc:
>> >>>>>>
>> >>>>>> http://www.linux-kvm.org/page/WindowsGuestDrivers/kvmnet/registry
>> >>>>>>
>> >>>>>> and upgraded my virtio network driver, but it's not helped.  I've
>> also
>> >>>>>> done about an hour or two of googling which has revealed little.
>> >>>>>>
>> >>>>>> I understand that this is not an openstack issue, but, I suspect,
>> >> others
>> >>>>>> have encountered this when bringing 2008 R2 guests into their
>> >> clusters.
>> >>>>>> Anyone?
>> >>>>>>
>> >>>>>> Thanks much,
>> >>>>>> JR
>> >>>>>>
>> >>>>>>
>> >>>>>>
>> >>>>>> _______________________________________________
>> >>>>>> Mailing list:
>> >>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
>> >>>>>> Post to     : openstack at lists.openstack.org
>> >>>>>> Unsubscribe :
>> >>>>>> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack
>> >>>>>>
>> >>>>>
>> >>>>
>> >>>> --
>> >>>> Your electronic communications are being monitored; strong
>> encryption is
>> >>>> an answer. My public key
>> >>>> <
>> http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x4F08C504BD634953>
>> >>>>
>> >>>
>> >>
>> >> --
>> >> Your electronic communications are being monitored; strong encryption
>> is
>> >> an answer. My public key
>> >> <http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x4F08C504BD634953>
>> >>
>> >
>>
>> --
>> Your electronic communications are being monitored; strong encryption is
>> an answer. My public key
>> <http://pgp.mit.edu:11371/pks/lookup?op=get&search=0x4F08C504BD634953>
>>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.openstack.org/pipermail/openstack/attachments/20140509/772becc2/attachment.html>


More information about the Openstack mailing list