[lxc-devel] lxc final thesis

[Farcasi Ana-Maria]

Hello,

As mentioned in our previous email, we were having issues getting bandwidth
limitation working on cgroup ( with net_cls ) and tc. We've been running
various tests and scenarios but found no way to enable proper limitation.
Throughout our testing we've created a virtual machine running a 2.6.35
kernel where the limitation was indeed working.

We've browsed recent commits in the Linux-2.6 kernel source code and
discovered that ever since version 2.6.35 (actually ever since this
commit[1] was integrated), there is a change in getting the packet classid -
the classid is stored and read as a member of a struct sock. As this is also
used by LXC, it means that ever since 2.6.35, bandwidth limitation using tc
is working for cgroups.

In order to prove this, we've compiled the latest 2.6.34 and 2.6.35 kernels
(2.6.34.8 and 2.6.35.11). We've used tc, cgroup (net_cls) and iperf to test
this.

On the latter version ( 2.6.35.11 ) the limitation is working accordingly,
while on the former ( 2.6.34.8 ) the limitation is not working at all. We
believe the above mentioned commit[1] is responsible for solving this issue
and post-2.6.35 kernels should have no problems using cgroup-based bandwidth
limitation (for example, within LXC).

We were thinking whether it would be a good idea to integrate an option for
bandwidth limitation into an LXC container configuration file (such as
lxc.network.bwlimit). This would allow a rapid setup of an LXC container and
network limitation. This could, of course, be set up using tc (it would take
a bit more effort, though). What do you think?

Thank you,
Ana & Irina.

[1]
http://git.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commit;h=f845172531fb7410c7fb7780b1a6e51ee6df7d52
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.linuxcontainers.org/pipermail/lxc-devel/attachments/20110318/cda42d8a/attachment.html>