Now for part two in my ongoing series on making Xen suck less. Last time we looked at making networking work for hardware virtualized machines. Networking for paravirtualized VMs does work out of the box, but this hint might help if you’re running into performance problems.
Paravirtualized Networking Performance
If you’re running a web server or some other server that’s sending large files (or sometimes small files), you may find that your VM seems to hang inexplicably on those transfers.
For some reason, the paravirtualized Xen networking drivers advertise that they support on-board TCP segmentation. In fact, they seem to pass the packets onto the wire un-segmented, which frequently will cause the packets to be dropped for going over the MTU.
If you’re using xen-create-image, there’s a commented out line in /etc/network/interfaces that runs ethtool -K eth0 tx off. That’s close to the right issue. You actually want to add a line to your /etc/network/interfaces so that it looks something like this:
auto eth0 iface eth0 inet static address 18.181.0.80 gateway 18.181.0.1 netmask 255.255.0.0 post-up ethtool -K eth0 tso off
Stay tuned for more hints, including how to deal with clock issues and magic sysrqs. I’ll also be pulling walkthroughs together on converting paravirtualized to hardware virtualized VMs, and how to upgrade older Ubuntu releases to more recent ones safely.