[Qemu-devel] Host networking hang-ups in a qemu cluster

[Qemu-devel] Host networking hang-ups in a qemu cluster

[Kfir Lavi]

[-- Attachment #1: Type: text/plain, Size: 1421 bytes --]

Hi,
I have a cluster of similar qemu machines, for doing some simulation.
I'm using e1000 as network driver.
When loading few machines, we encounter hangs of networking inside a
virtual machine, and
it will hang all the interfaces inside the machine.
If I'll do down/up to all interfaces, the networking will resume.
Reboot of the vm will solve the problem the same way.
Please not that there are some bursts of packets, but really not a lot.
To mitigate burst problems, I connected each tap to a bridge, and it seems
to
have some effect. (each tap has different vlan, so it gets only its
traffic).
I tried to use ethtool rx tx as 4096 in the virtual machines.
Virtio cause more problems then e1000.
What is your take on it, where should I check? Is there any way to debug
the
buffers of qemu?
Each vm has 6 interfaces connected to tap device on the host.
Is it possible that qemu networking buffer is shared between all the vm
interfaces?
A software we are running inside the virtual machines opens a raw socket
(read, write) to intercept packages.
Is it possible that this raw socket can hang all the interfaces in the
machine?

The virtual machines are running Gentoo with 2.6.28 kernel (Embedded), and
the qemu host running
Gentoo with kernel 3.1.1 with nested kvm and unstable (recent) qemu.

Can it be the mismatch of the 2.6.28 in the vm, and the recent qemu and
latest kvm from 3.1.1?

Thanks for your help,
Kfir

[-- Attachment #2: Type: text/html, Size: 1565 bytes --]