[Openstack] URGENT: packet loss on openstack instance

Satish Patel satish.txt at gmail.com
Sun Sep 16 17:41:29 UTC 2018


I successful reproduce this error with hping3 tool and look like
multiqueue is our solution :) but i have few question you may have
answer of that.

1. I have created two instance  (vm1.example.com & vm2.example.com)

2. I have flood traffic from vm1 using "hping3 vm2.example.com
--flood"  and i have noticed drops on tap interface. ( This is without
multiqueue)

3. Enable multiqueue in image and run same test and again got packet
drops on tap interface ( I didn't update queue on vm2 guest, so
definitely i was expecting packet drops)

4. Now i have try to update vm2 queue using ethtool and i got
following error, I have 15vCPU and i was trying to add 15 queue

[root at bar-mq ~]# ethtool -L eth0 combined 15
Cannot set device channel parameters: Invalid argument

Then i have tried 8 queue which works.

[root at bar-mq ~]# ethtool -L eth0 combined 8
combined unmodified, ignoring
no channel parameters changed, aborting
current values: tx 0 rx 0 other 0 combined 8

Now i am not seeing any packet drops on tap interface, I have measure
PPS and i was able to get 160kpps without packet drops.

Question:

1. why i am not able to add 15 queue?  ( is this NIC or driver limitation?)
2. how do i automate "ethtool -L eth0 combined 8" command in instance
so i don't need to tell my customer to do this manually?
On Sun, Sep 16, 2018 at 11:53 AM Satish Patel <satish.txt at gmail.com> wrote:
>
> Hi Liping,
>
> >> I think multi queue feature should help.(be careful to make sure the ethtool update queue number action also did after reboot the vm).
>
> Is there a way i can automate this last task to update queue number
> action after reboot vm :) otherwise i can use cloud-init to make sure
> all VM build with same config.
> On Sun, Sep 16, 2018 at 11:51 AM Satish Patel <satish.txt at gmail.com> wrote:
> >
> > I am currently playing with those setting and trying to generate
> > traffic with hping3 tools, do you have any tool to test traffic
> > performance for specially udp style small packets.
> >
> > I am going to share all my result and see what do you feel because i
> > have noticed you went through this pain :)  I will try every single
> > option which you suggested to make sure we are good before i move
> > forward to production.
> > On Sun, Sep 16, 2018 at 11:25 AM Liping Mao (limao) <limao at cisco.com> wrote:
> > >
> > > I think multi queue feature should help.(be careful to make sure the ethtool update queue number action also did after reboot the vm).
> > >
> > > Numa cpu pin and queue length will be a plus in my exp. You may need yo have performance test in your situatuon,in my case cpu numa helpped the app get very stable 720p/1080p transcoding performance. Not sure if your app get benifit.
> > >
> > > You are not using L3,this will let you avoid a lot of performance issue. And since only two instance with 80kpps packets,so in your case,HW interface should not be bottleneck too. And your Nexus 5k/7k will not be bottleneck for sure ;-)
> > >
> > >
> > > Thanks,
> > > Liping Mao
> > >
> > > > 在 2018年9月16日,23:09,Satish Patel <satish.txt at gmail.com> 写道:
> > > >
> > > > Thanks Liping,
> > > >
> > > > I am using libvertd 3.9.0 version so look like i am eligible take
> > > > advantage of that feature. phew!
> > > >
> > > > [root at compute-47 ~]# libvirtd -V
> > > > libvirtd (libvirt) 3.9.0
> > > >
> > > > Let me tell you how i am running instance on my openstack, my compute
> > > > has 32 core / 32G memory  and i have created two instance on compute
> > > > node 15vcpu and 14G memory ( two instance using 30 vcpu core, i have
> > > > kept 2 core for compute node). on compute node i disabled overcommit
> > > > using ratio (1.0)
> > > >
> > > > I didn't configure NUMA yet because i wasn't aware of this feature, as
> > > > per your last post do you think numa will help to fix this issue?
> > > > following is my numa view
> > > >
> > > > [root at compute-47 ~]# numactl --hardware
> > > > available: 2 nodes (0-1)
> > > > node 0 cpus: 0 1 2 3 4 5 6 7 16 17 18 19 20 21 22 23
> > > > node 0 size: 16349 MB
> > > > node 0 free: 133 MB
> > > > node 1 cpus: 8 9 10 11 12 13 14 15 24 25 26 27 28 29 30 31
> > > > node 1 size: 16383 MB
> > > > node 1 free: 317 MB
> > > > node distances:
> > > > node   0   1
> > > >  0:  10  20
> > > >  1:  20  10
> > > >
> > > >
> > > > I am not using any L3 router, i am using provide VLAN network and
> > > > using Cisco Nexus switch for my L3 function so i am not seeing any
> > > > bottleneck there.
> > > >
> > > > This is the 10G NIC i have on all my compute node, dual 10G port with
> > > > bonding (20G)
> > > >
> > > > 03:00.0 Ethernet controller: Broadcom Limited NetXtreme II BCM57810 10
> > > > Gigabit Ethernet (rev 10)
> > > > 03:00.1 Ethernet controller: Broadcom Limited NetXtreme II BCM57810 10
> > > > Gigabit Ethernet (rev 10)
> > > >
> > > >
> > > >> On Sun, Sep 16, 2018 at 10:50 AM Liping Mao (limao) <limao at cisco.com> wrote:
> > > >>
> > > >> It is still possible to update rx and tx queues length if your qemu and libvirt version is higher than the version recorded in [3]. (You should possible to update directly in libvirt configuration if my memory is correct)
> > > >>
> > > >> We also have some similar use case which run audio/vedio serivcs. They are CPU consuming and have UDP small packets. Another possible tunning is using CPU pin for the vm.  you can use numa awared cpu feature to get stable cpu performance ,vm network dropped packets sometimes because of the vm cpu is too busy,with numa cpu it works better performance,our way is similar with [a]. You need to create flavor with special metadata and dedicated Host Agg for numa awared VMs. Dedicated CPU is very good for media service. It makes the CPU performance stable.
> > > >>
> > > >> Another packet loss case we get is because of vm kernel, some of our app are using 32bit OS, that cause memory issue, when traffic larger then 50kpps, it dropped a lot,sometimes,it even crash. In this case, 32bit os can actually use very limited memory, we have to add swap for the vm. Hope your app is using 64 bit OS. Because 32 bit could cause tons of trouble.
> > > >>
> > > >> BTW,if you are using vrouter on L3, you’d better to move provider network(no vrouter). I did not tried DVR, but if you are running without DVR, the L3 node will be bottleneck very quick. Especially default iptables conntrack is 65535, you will reach to it and drop packet on L3, even after you tun that value, it still hard to more that 1Mpps for your network node.
> > > >>
> > > >> If your App more than 200kpps per compute node, you may be better also have a look your physical network driver tx/rx configuration. Most of the HW default value for tx/rx queues number and length are very poor,you may start to get packet on eth interface on physical host when rx queue is full.
> > > >>
> > > >> [a]https://redhatstackblog.redhat.com/2015/05/05/cpu-pinning-and-numa-topology-awareness-in-openstack-compute/
> > > >>
> > > >> Regards,
> > > >> Liping Mao
> > > >>
> > > >> 在 2018年9月16日,21:18,Satish Patel <satish.txt at gmail.com> 写道:
> > > >>
> > > >> Hi Liping,
> > > >>
> > > >> Thank you for your reply,
> > > >>
> > > >> We notice packet drops during high load, I did try txqueue and didn't help so I believe I am going to try miltiqueue.
> > > >>
> > > >> For SRIOV I have to look if I have support in my nic.
> > > >>
> > > >> We are using queens so I think queue size option  not possible :(
> > > >>
> > > >> We are using voip application and traffic is udp so our pps rate is 60k to 80k per vm instance.
> > > >>
> > > >> I will share my result as soon as I try multiqueue.
> > > >>
> > > >>
> > > >>
> > > >> Sent from my iPhone
> > > >>
> > > >> On Sep 16, 2018, at 2:27 AM, Liping Mao (limao) <limao at cisco.com> wrote:
> > > >>
> > > >>
> > > >> Hi Satish,
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> Did your packet loss happen always or it only happened when heavy load?
> > > >>
> > > >>
> > > >> AFAIK, if you do not tun anything, the vm tap can process about 50kpps before the tap device start to drop packets.
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> If it happened in heavy load, couple of things you can try:
> > > >>
> > > >>
> > > >> 1) increase tap queue length, usually the default value is 500, you can try larger. (seems like you already tried)
> > > >>
> > > >>
> > > >> 2) Try to use virtio multi queues feature , see [1]. Virtio use one queue for rx/tx in vm, with this feature you can get more queues. You can check
> > > >>
> > > >>
> > > >> 3) In rock version, you can use [2] to increase virtio queue size, the default queues size is 256/512, you may increase it to 1024, this would help to increase pps of the tap device.
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> If all these things can not get your network performance requirement, you may need to move to use dpdk / sriov stuff to get more vm performance.
> > > >>
> > > >>
> > > >> I did not actually used them in our env, you may refer to [3]
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> [1] https://specs.openstack.org/openstack/nova-specs/specs/liberty/implemented/libvirt-virtiomq.html
> > > >>
> > > >>
> > > >> [2] https://specs.openstack.org/openstack/nova-specs/specs/rocky/implemented/libvirt-virtio-set-queue-sizes.html
> > > >>
> > > >>
> > > >> [3] https://docs.openstack.org/ocata/networking-guide/config-sriov.html
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> Regards,
> > > >>
> > > >>
> > > >> Liping Mao
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> 在 2018/9/16 13:07,“Satish Patel”<satish.txt at gmail.com> 写入:
> > > >>
> > > >>
> > > >>
> > > >>
> > > >>  [root at compute-33 ~]# ifconfig tap5af7f525-5f | grep -i drop
> > > >>
> > > >>
> > > >>          RX errors 0 dropped 0 overruns 0 frame 0
> > > >>
> > > >>
> > > >>          TX errors 0 dropped 2528788837 overruns 0 carrier 0 collisions 0
> > > >>
> > > >>
> > > >>
> > > >>
> > > >>  Noticed tap interface dropping TX packets and even after increasing
> > > >>
> > > >>
> > > >>  txqueue from 1000 to 10000 nothing changed, still getting packet
> > > >>
> > > >>
> > > >>  drops.
> > > >>
> > > >>
> > > >>
> > > >>
> > > >>  On Sat, Sep 15, 2018 at 4:22 PM Satish Patel <satish.txt at gmail.com> wrote:
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> Folks,
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> I need some advice or suggestion to find out what is going on with my
> > > >>
> > > >>
> > > >> network, we have notice high packet loss on openstack instance and not
> > > >>
> > > >>
> > > >> sure what is going on, same time if i check on host machine and it has
> > > >>
> > > >>
> > > >> zero packet loss.. this is what i did for test...
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> ping 8.8.8.8
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> from instance: 50% packet loss
> > > >>
> > > >>
> > > >> from compute host: 0% packet loss
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> I have disabled TSO/GSO/SG setting on physical compute node but still
> > > >>
> > > >>
> > > >> getting packet loss.
> > > >>
> > > >>
> > > >>
> > > >>
> > > >> We have 10G NIC on our network, look like something related to tap
> > > >>
> > > >>
> > > >> interface setting..
> > > >>
> > > >>
> > > >>
> > > >>
> > > >>  _______________________________________________
> > > >>
> > > >>
> > > >>  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
> > > >>
> > > >>
> > > >>
> > > >>
> > > >>



More information about the Openstack mailing list