2

We are running a number of Debian Wheezy VMs on top of Ubuntu Servers 12.04.4 / Libvirt 0.9.8-2ubuntu17.17

The host is connected to the network through a trunk. It then split the VLANs and create a bridge for each of them with the following conf:

auto eth4.2 kvmbrtrunk.2
iface eth4.2 inet manual
    up ifconfig eth4.2 up
    down ifconfig eth4.2 down

iface kvmbrtrunk.2 inet manual
    bridge-ports eth4.2
    bridge-stp no
    bridge-fd 0
    bridge-maxwait 0

The VMs are configured as follow:

 <interface type='bridge'>
   <mac address='54:52:00:02:10:70'/>
   <source bridge='kvmbrtrunk.2'/>
   <model type='virtio'/>
   <address type='pci' domain='0x0000' bus='0x00' slot='0x09' function='0x0'/>
 </interface>

And they use VirtIO

00:09.0 Ethernet controller: Red Hat, Inc Virtio network device

Protagonists (all on same VLAN):

A: 1st Ubuntu 12.04 desktop
B: 2nd Ubuntu 12.04 desktop
C: 1st VM, 1st host
D: 2nd VM, 1st host
E: 3rd VM, 2nd host

When we do a serie of 60 pings "rtt min/avg/max/mdev":

A -> B = 0.093/0.132/0.158/0.015 ms
A -> C = 0.272/0.434/1.074/0.113 ms
A -> D = 0.294/0.460/0.832/0.091 ms
A -> E = 0.324/0.505/0.831/0.069 ms
C -> D = 0.348/0.607/0.863/0.124 ms
C -> E = 0.541/0.792/0.972/0.101 ms

So those results seem to indicate that the libvirt's virtual switch/filtering not only adds some latency, as one might expect, but triples it (0.132 vs 0.460)

Question

 Is there anything that can be done to attenuate this extra latency?

Thanks in advance for any tips.

Jef Bucas
  • 21
  • 2

1 Answers1

1

What kind of features are you willing to sacrifice for this reduction in latency?

To start out with, try disabling iptables/ebtables on the bridge interfaces. You can change /proc/sys/net/bridge/bridge-nf-call-iptables to 0 to accomplish this. The downside here is you can no longer do any sort of filtering on the guest traffic.

The 'better' option here is to switch to virtualized network cards, using SR-IOV. This requires that your motherboard and network controller support it. You're also limited to 7 guests per network card (for gigabit ethernet, I believe 10gig is higher). This works by giving each guest direct access to the networking hardware. The hosts OS is not involved in the packet flow, the VM just talks directly to the NIC.

SR-IOV will give you the best performance (CPU usage was around 10% lower in our tests, for high levels of network traffc), and lowest latency (as there are far fewer layers of software interacting with the packets). I believe you can configure vlan tagging with this, but the setup would likely be a bit difficult (SR-IOV is basically undocumented magic, and you'll be doing a lot of fiddling with settings).

devicenull
  • 5,572
  • 1
  • 25
  • 31