Hi Michal,

these are the steps:

- start the vm with qemu

sudo qemu-system-x86_64-spice -m 2048 -enable-kvm -smp 3 -cdrom /home/machine_A/ubuntu.iso -netdev tap,ifname=tap0,vhost=on,id=n1,vhostforce=on,queues=3,script=/etc/tap0_ifup.sh,downscript=/etc/tap0_ifdown.sh -device virtio-net-pci,netdev=n1,mac=9E:58:00:d2:53:03,mq=on,vectors=8 -netdev tap,ifname=tap1,vhost=on,id=n2,vhostforce=on,queues=3,script=/etc/tap1_ifup.sh,downscript=/etc/tap1_ifdown.sh -device virtio-net-pci,netdev=n2,mac=7A:53:00:d1:59:04,mq=on,vectors=8 -device virtio-net,netdev=network2 -netdev user,id=network2 -hda switch_A.img

- attach the tap to the bridges in the host machine (to have the traffic coming from the outsider traffic generator injected in the vm)

- in the VM:

- enable ip_forward

- enable promiscuous mode for the interface

- send the "tc qdisc" command to configure the output NIC with 2 traffic classes (TC) assigned to 2 different queues

- add the iptables mangle rules to the POSTROUTING chain to assign TC1 to the traffic with dport 7777 and TC0 to the traffic with dport 8888

I'm sending 2 UDP flows, I have no loss under 3Mbps each.

If I capture traffic with tcpdump on the ingress and on the egress nic (of the VM), I see a difference of 50% of packets

When I did a test on the host machine, to check if it has the same problem, I've reached an aggregated traffic of 90Mbps with no loss.

That's why I think that there is some misconfiguration on the virtual NIC.

Thanks

Silvia

On Mon, Mar 29, 2021 at 11:13 AM Michal Privoznik <mprivozn@redhat.com> wrote:

On 3/27/21 2:53 PM, Silvia Fichera wrote:
> Hi all,
> I want to use tc qdisc settings in a network coposed of several qemu
> VMs, connected through bridges and tap interfaces.
> I generate traffic with a spirent. Everything is fine when the
> scheduling discipline is not installed but when I run the command to set
> taprio queues traffic on the VM's NIC the traffic is dropped, i can send
> max 1mbps.
> I think that there is something missing in the virtual NIC configuration
> or setup. With ethtool i can see that queues are configured. I've also
> noticed the BQL equals to 0, that is different than the physical machine
> (BQL=18600) where everything works correctly.
> I've read that it could be because NIC drivers do not support that setting.
>
> Do you have any suggestions?

Hey,

I'm not familiar with taprio, but what's implemented in libvirt is htb
and sfq and that works well. Are you setting qdisc-s yourself or
modifying libvirt created structure?

Are you setting these qdiscs from the host, right?
I know that when changing QoS settings (when libvirt changes qdics/class
layout) for a brief moment packets are not transmitted from/to guest. I
suspect that kernel is freeing up queues or something. But this does not
look like your case, does it?

Michal