[openstack][neutron][openvswitch] Openvswitch Packet loss when high throughput (pps)

Satish Patel satish.txt at gmail.com
Fri Sep 8 01:57:36 UTC 2023


I would say let's run your same benchmark with OVS-DPDK and tell me if you
see better performance. I doubt you will see significant performance boot
but lets see. Please prove me wrong :)

On Thu, Sep 7, 2023 at 9:45 PM Ha Noi <hanoi952022 at gmail.com> wrote:

> Hi Satish,
>
> Actually, the guess interface is not using tap anymore.
>
>     <interface type='vhostuser'>
>       <mac address='fa:16:3e:76:77:dd'/>
>       <source type='unix' path='/var/run/openvswitch/vhu3766ee8a-86'
> mode='server'/>
>       <target dev='vhu3766ee8a-86'/>
>       <model type='virtio'/>
>       <alias name='net0'/>
>       <address type='pci' domain='0x0000' bus='0x00' slot='0x03'
> function='0x0'/>
>     </interface>
>
> It's totally bypass the kernel stack ?
>
>
>
>
> On Fri, Sep 8, 2023 at 5:02 AM Satish Patel <satish.txt at gmail.com> wrote:
>
>> I did test OVS-DPDK and it helps offload the packet process on compute
>> nodes, But what about VMs it will still use a tap interface to attach from
>> compute to vm and bottleneck will be in vm. I strongly believe that we have
>> to run DPDK based guest to pass through the kernel stack.
>>
>> I love to hear from other people if I am missing something here.
>>
>> On Thu, Sep 7, 2023 at 5:27 PM Ha Noi <hanoi952022 at gmail.com> wrote:
>>
>>> Oh. I heard from someone on the reddit said that Ovs-dpdk is transparent
>>> with user?
>>>
>>> So It’s not correct?
>>>
>>> On Thu, 7 Sep 2023 at 22:13 Satish Patel <satish.txt at gmail.com> wrote:
>>>
>>>> Because DPDK required DPDK support inside guest VM. It's not
>>>> suitable for general purpose workload. You need your guest VM network to
>>>> support DPDK to get 100% throughput.
>>>>
>>>> On Thu, Sep 7, 2023 at 8:06 AM Ha Noi <hanoi952022 at gmail.com> wrote:
>>>>
>>>>> Hi Satish,
>>>>>
>>>>> Why dont you use DPDK?
>>>>>
>>>>> Thanks
>>>>>
>>>>> On Thu, 7 Sep 2023 at 19:03 Satish Patel <satish.txt at gmail.com> wrote:
>>>>>
>>>>>> I totally agreed with Sean on all his points but trust me, I have
>>>>>> tried everything possible to tune OS, Network stack, multi-queue, NUMA, CPU
>>>>>> pinning and name it.. but I didn't get any significant improvement. You may
>>>>>> gain 2 to 5% gain with all those tweek. I am running the entire workload on
>>>>>> sriov and life is happy except no LACP bonding.
>>>>>>
>>>>>> I am very interesting is this project
>>>>>> https://docs.openvswitch.org/en/latest/intro/install/afxdp/
>>>>>>
>>>>>> On Thu, Sep 7, 2023 at 6:07 AM Ha Noi <hanoi952022 at gmail.com> wrote:
>>>>>>
>>>>>>> Dear Smoney,
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>> On Thu, Sep 7, 2023 at 12:41 AM <smooney at redhat.com> wrote:
>>>>>>>
>>>>>>>> On Wed, 2023-09-06 at 11:43 -0400, Satish Patel wrote:
>>>>>>>> > Damn! We have noticed the same issue around 40k to 55k PPS. Trust
>>>>>>>> me
>>>>>>>> > nothing is wrong in your config. This is just a limitation of the
>>>>>>>> software
>>>>>>>> > stack and kernel itself.
>>>>>>>> its partly determined by your cpu frequency.
>>>>>>>> kernel ovs of yesteryear could handel about 1mpps total on a ~4GHZ
>>>>>>>> cpu. with per port troughpuyt being lower dependin on what
>>>>>>>> qos/firewall
>>>>>>>> rules that were apllied.
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>> My CPU frequency is 3Ghz and using CPU Intel Gold 2nd generation. I
>>>>>>> think the problem is tuning in the compute node inside. But I cannot find
>>>>>>> any guide or best practices for it.
>>>>>>>
>>>>>>>
>>>>>>>
>>>>>>>> moving form iptables firewall to ovs firewall can help to some
>>>>>>>> degree
>>>>>>>> but your partly trading connection setup time for statead state
>>>>>>>> troughput
>>>>>>>> with the overhead of the connection tracker in ovs.
>>>>>>>>
>>>>>>>> using stateless security groups can help
>>>>>>>>
>>>>>>>> we also recently fixed a regression cause by changes in newer
>>>>>>>> versions of ovs.
>>>>>>>> this was notable in goign form rhel 8 to rhel 9 where litrally it
>>>>>>>> reduced
>>>>>>>> small packet performce to 1/10th and jumboframes to about 1/2
>>>>>>>> on master we have a config option that will set the default qos on
>>>>>>>> a port to linux-noop
>>>>>>>>
>>>>>>>> https://github.com/openstack/os-vif/blob/master/vif_plug_ovs/ovs.py#L106-L125
>>>>>>>>
>>>>>>>> the backports are propsoed upstream
>>>>>>>> https://review.opendev.org/q/Id9ef7074634a0f23d67a4401fa8fca363b51bb43
>>>>>>>> and we have backported this downstream to adress that performance
>>>>>>>> regression.
>>>>>>>> the upstram backport is semi stalled just ebcasue we wanted to
>>>>>>>> disucss if we shoudl make ti opt in
>>>>>>>> by default upstream while backporting but it might be helpful for
>>>>>>>> you if this is related to yoru current
>>>>>>>> issues.
>>>>>>>>
>>>>>>>> 40-55 kpps is kind of low for kernel ovs but if you have a low
>>>>>>>> clockrate cpu, hybrid_plug + incorrect qos
>>>>>>>> then i could see you hitting such a bottelneck.
>>>>>>>>
>>>>>>>> one workaround by the way without the os-vif workaround backported
>>>>>>>> is to set
>>>>>>>> /proc/sys/net/core/default_qdisc to not apply any qos or a low
>>>>>>>> overhead qos type
>>>>>>>> i.e. sudo sysctl -w net.core.default_qdisc=pfifo_fast
>>>>>>>>
>>>>>>>>
>>>>>>>
>>>>>>>> that may or may not help but i would ensure that your are not usign
>>>>>>>> somting like fqdel or cake
>>>>>>>> for net.core.default_qdisc and if you are try changing it to
>>>>>>>> pfifo_fast and see if that helps.
>>>>>>>>
>>>>>>>> there isnet much you can do about the cpu clock rate but ^ is
>>>>>>>> somethign you can try for free
>>>>>>>> note it wont actully take effect on an exsitng vm if you jsut
>>>>>>>> change the default but you can use
>>>>>>>> tc to also chagne the qdisk for testing. hard rebooting the vm
>>>>>>>> shoudl also make the default take effect.
>>>>>>>>
>>>>>>>> the only other advice i can give assuming kernel ovs is the only
>>>>>>>> option you have is
>>>>>>>>
>>>>>>>> to look at
>>>>>>>>
>>>>>>>> https://docs.openstack.org/nova/latest/configuration/config.html#libvirt.rx_queue_size
>>>>>>>>
>>>>>>>> https://docs.openstack.org/nova/latest/configuration/config.html#libvirt.tx_queue_size
>>>>>>>> and
>>>>>>>>
>>>>>>>> https://docs.openstack.org/nova/latest/configuration/extra-specs.html#hw:vif_multiqueue_enabled
>>>>>>>>
>>>>>>>> if the bottelneck is actully in qemu or the guest kernel rather
>>>>>>>> then ovs adjusting the rx/tx queue size and
>>>>>>>> using multi queue can help. it will have no effect if ovs is the
>>>>>>>> bottel neck.
>>>>>>>>
>>>>>>>>
>>>>>>>>
>>>>>>> I have set this option to 1024, and enable multiqueue as well. But
>>>>>>> it did not help.
>>>>>>>
>>>>>>>
>>>>>>>> >
>>>>>>>> > On Wed, Sep 6, 2023 at 9:21 AM Ha Noi <hanoi952022 at gmail.com>
>>>>>>>> wrote:
>>>>>>>> >
>>>>>>>> > > Hi Satish,
>>>>>>>> > >
>>>>>>>> > > Actually, our customer get this issue when the tx/rx above only
>>>>>>>> 40k pps.
>>>>>>>> > > So what is the threshold of this throughput for OvS?
>>>>>>>> > >
>>>>>>>> > >
>>>>>>>> > > Thanks and regards
>>>>>>>> > >
>>>>>>>> > > On Wed, 6 Sep 2023 at 20:19 Satish Patel <satish.txt at gmail.com>
>>>>>>>> wrote:
>>>>>>>> > >
>>>>>>>> > > > Hi,
>>>>>>>> > > >
>>>>>>>> > > > This is normal because OVS or LinuxBridge wire up VMs using
>>>>>>>> TAP interface
>>>>>>>> > > > which runs on kernel space and that drives higher interrupt
>>>>>>>> and that makes
>>>>>>>> > > > the kernel so busy working on handling packets. Standard
>>>>>>>> OVS/LinuxBridge
>>>>>>>> > > > are not meant for higher PPS.
>>>>>>>> > > >
>>>>>>>> > > > If you want to handle higher PPS then look for DPDK or SRIOV
>>>>>>>> deployment.
>>>>>>>> > > > ( We are running everything in SRIOV because of high PPS
>>>>>>>> requirement)
>>>>>>>> > > >
>>>>>>>> > > > On Tue, Sep 5, 2023 at 11:11 AM Ha Noi <hanoi952022 at gmail.com>
>>>>>>>> wrote:
>>>>>>>> > > >
>>>>>>>> > > > > Hi everyone,
>>>>>>>> > > > >
>>>>>>>> > > > > I'm using Openstack Train and Openvswitch for ML2 driver
>>>>>>>> and GRE for
>>>>>>>> > > > > tunnel type. I tested our network performance between two
>>>>>>>> VMs and suffer
>>>>>>>> > > > > packet loss as below.
>>>>>>>> > > > >
>>>>>>>> > > > > VM1: IP: 10.20.1.206
>>>>>>>> > > > >
>>>>>>>> > > > > VM2: IP: 10.20.1.154 <https://10.20.1.154/24>
>>>>>>>> > > > >
>>>>>>>> > > > > VM3: IP: 10.20.1.72
>>>>>>>> > > > >
>>>>>>>> > > > >
>>>>>>>> > > > > Using iperf3 to testing performance between VM1 and VM2.
>>>>>>>> > > > >
>>>>>>>> > > > > Run iperf3 client and server on both VMs.
>>>>>>>> > > > >
>>>>>>>> > > > > On VM2: iperf3 -t 10000 -b 130M -l 442 -P 6 -u -c
>>>>>>>> 10.20.1.206
>>>>>>>> > > > >
>>>>>>>> > > > > On VM1: iperf3 -t 10000 -b 130M -l 442 -P 6 -u -c
>>>>>>>> 10.20.1.154
>>>>>>>> > > > > <https://10.20.1.154/24>
>>>>>>>> > > > >
>>>>>>>> > > > >
>>>>>>>> > > > > Using VM3 ping into VM1, then the packet is lost and the
>>>>>>>> latency is
>>>>>>>> > > > > quite high.
>>>>>>>> > > > >
>>>>>>>> > > > >
>>>>>>>> > > > > ping -i 0.1 10.20.1.206
>>>>>>>> > > > >
>>>>>>>> > > > > PING 10.20.1.206 (10.20.1.206) 56(84) bytes of data.
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=1 ttl=64 time=7.70 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=2 ttl=64 time=6.90 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=3 ttl=64 time=7.71 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=4 ttl=64 time=7.98 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=6 ttl=64 time=8.58 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=7 ttl=64 time=8.34 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=8 ttl=64 time=8.09 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=10 ttl=64 time=4.57 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=11 ttl=64 time=8.74 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=12 ttl=64 time=9.37 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=14 ttl=64 time=9.59 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=15 ttl=64 time=7.97 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=16 ttl=64 time=8.72 ms
>>>>>>>> > > > >
>>>>>>>> > > > > 64 bytes from 10.20.1.206: icmp_seq=17 ttl=64 time=9.23 ms
>>>>>>>> > > > >
>>>>>>>> > > > > ^C
>>>>>>>> > > > >
>>>>>>>> > > > > --- 10.20.1.206 ping statistics ---
>>>>>>>> > > > >
>>>>>>>> > > > > 34 packets transmitted, 28 received, 17.6471% packet loss,
>>>>>>>> time 3328ms
>>>>>>>> > > > >
>>>>>>>> > > > > rtt min/avg/max/mdev = 1.396/6.266/9.590/2.805 ms
>>>>>>>> > > > >
>>>>>>>> > > > >
>>>>>>>> > > > >
>>>>>>>> > > > > Does any one get this issue ?
>>>>>>>> > > > >
>>>>>>>> > > > > Please help me. Thanks
>>>>>>>> > > > >
>>>>>>>> > > >
>>>>>>>>
>>>>>>>>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.openstack.org/pipermail/openstack-discuss/attachments/20230907/74e830be/attachment-0001.htm>


More information about the openstack-discuss mailing list