[openstack-dev] [neutron][rootwrap] Performance considerations, sudo?
carl at ecbaldwin.net
Fri Mar 7 17:47:00 UTC 2014
I had a reply drafted up to Miguel's original post and now I realize
that I never actually sent it. :( So, I'll clean up and update my
draft and send it. This is a huge impediment to scaling Neutron and I
believe this needs some attention before Icehouse releases.
I believe this problem needs to be tackled on multiple fronts. I have
been focusing mostly on the L3 agent because routers seem to take a
lot more commands to create and maintain than DHCP namespaces, in
general. I've created a few patches to address the issues that I've
found. The patch that Mark mentioned  is one potential part of the
solution but it turns out to be one of the more complicated patches to
work out and it keeps falling lower in priority for me. I have come
back to it this week and will work on it through next week as a higher
There are some other recent improvements that have merged to Icehouse
3: I have changed the iptables lock to avoid contention , avoided
an unnecessary RPC call for each router processed , and avoided
some unnecessary ip netns calls to check existence of a device . I
feel like I'm just slowly whittling away at the problem.
I'm also throwing around the idea of refactoring the L3 agent to give
precedence to RPC calls on a restart . There is a very rough
preview up that I put up yesterday evening to get feedback on the
approach that I'm thinking of taking. This should make the agent more
responsive to changes that come in through RPC. This is less of a win
on reboot than on a simple agent process restart.
Another thing that we've found to help is to delete namespaces when a
router or dhcp server namespace is no longer needed . We've
learned that having vestigial namespaces hanging around and
accumulating when they are no longer needed adversely affects the
performance of all "ip netns exec" commands. There are some sticky
kernel issues related to using this patch. That is why the default
configuration is to not delete namespaces. See the "Related-Bug"
referenced by that commit message.
I'm intrigued by the idea of writing a rootwrap compatible alternative
in C. It might even be possible to replace sudo + rootwrap
combination with a single, stand-alone executable with setuid
capability of elevating permissions on its own. I know it breaks the
everything-in-python pattern that has been established but this sort
of thing is sensitive enough to start-up time that it may be worth it.
I think we've shown that some of the OpenStack projects, namely Nova
and Neutron, run enough commands at scale that this performance really
matters. My plate is full enough that I cannot imagine taking on this
kind of task at this time. Does anyone have any interest in making
this a reality?
A C version of rootwrap could do some of the more common and simple
command verification and punt anything that fails to the python
version of rootwrap with an exec. That would ease the burden of
keeping it in sync and feature compatible with the python version and
allow python developers to continue developing root wrap in python.
On Fri, Mar 7, 2014 at 9:22 AM, Mark McClain <mmcclain at yahoo-inc.com> wrote:
> On Mar 6, 2014, at 3:31 AM, Miguel Angel Ajo <majopela at redhat.com> wrote:
>> Yes, one option could be to coalesce all calls that go into
>> a namespace into a shell script and run this in the
>> ootwrap > ip netns exec
>> But we might find a mechanism to determine if some of the steps failed, and what was the result / output, something like failing line + result code. I'm not sure if we rely on stdout/stderr results at any time.
> This is exactly one of the items Carl Baldwin has been investigating. Have you checked out his early work? 
>  https://review.openstack.org/#/c/67490/
> OpenStack-dev mailing list
> OpenStack-dev at lists.openstack.org
More information about the OpenStack-dev