[ovs-discuss] ovs-dpdk performance is not good
gowrishankar
gowrishankar.m at linux.vnet.ibm.com
Tue Jul 14 13:00:06 UTC 2015
What I realised while I too observed similar lower perf with dpdk was,
because of TSO.
In case of VM just launched with virtio-net driver, most of segmentation
offload
settings are turned on and notably TSO. Due to this, tcp could throttle
for larger
segment even more than mtu 1500 (i.e 1448 data). Where as, in case of
vhost-net
offloaded through vhost-user, it seems TSO is fixed as "off" and can not
be turned on.
This limits tcp segment length not to exceed mtu. This also means, DPDK
would gain
on lower packet sizes (as shown below).
I disabled TSO in LINUX BR and OVS (2.3.1) scenarios and we see
ovs-dpdk's winning
figures. But that need not be preferable way of launching guests on
vhost-net normally.
Here are my observations:
LINUX BR OVS (v2.3.1) 1xPMD 4xPMD 8xPMD
BW 6599 6367 2409 2270 1642
RR 1674 1830 13296 14680 12541
<AFTER DISABLING TSO IN VHOST-NET>
BW 1326 1507 2409 2270 1642
RR 1597 1830 13296 14680 12541
I had taken care of isolating cpus to run PMD threads as above. Earlier,
there was some
suggestions on other alternatives to do with MTU.
https://www.mail-archive.com/discuss@openvswitch.org/msg13912.html
It may help. Then I thought of making use of upstream multiqueue support
in dpdk but
it seems not yet supported yet in ovs. I am also curious to find
ovs-dpdk tuned for more
perf than ovs.
Regards,
Gowrishankar
On Tuesday 14 July 2015 05:55 PM, Traynor, Kevin wrote:
>
> *From:*discuss [mailto:discuss-bounces at openvswitch.org] *On Behalf Of
> *Na Zhu
> *Sent:* Monday, July 13, 2015 3:15 AM
> *To:* bugs at openvswitch.org
> *Subject:* [ovs-discuss] ovs-dpdk performance is not good
>
> Dear all,
>
> I want to use ovs-dpdk to improve my nfv performance. But when i
> compare the throughput between standard ovs and ovs-dpdk, the ovs is
> better, does anyone know why?
>
> I use netperf to test the throughput.
>
> use vhost-net to test standard ovs.
>
> use vhost-user to test ovs-dpdk.
>
> My topology is as follow:
>
> 内嵌图片 1
>
> The result is that standard ovs performance is better. Throughput unit
> Mbps.
>
> 内嵌图片 2
>
> 内嵌图片 3
>
> [kt] I would check your core affinitization to ensure that the vswitchd
>
> pmd is on a separate core to the vCPUs (set with
> other_config:pmd-cpu-mask).
>
> Also, this test is not using the DPDK vitrio PMD in the guest which
> provides
>
> performance gains.
>
> What packet sizes are you using? you should see a greater gain from DPDK
>
> at lower packet sizes (i.e. more PPS)
>
>
>
> _______________________________________________
> discuss mailing list
> discuss at openvswitch.org
> http://openvswitch.org/mailman/listinfo/discuss
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20150714/7438d171/attachment-0002.html>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: image/png
Size: 16645 bytes
Desc: not available
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20150714/7438d171/attachment-0006.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: image/png
Size: 11447 bytes
Desc: not available
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20150714/7438d171/attachment-0007.png>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: image/png
Size: 9044 bytes
Desc: not available
URL: <http://openvswitch.org/pipermail/ovs-discuss/attachments/20150714/7438d171/attachment-0008.png>
More information about the discuss
mailing list