Re: [iproute PATCH 0/2] Netns performance improvements

Phil Sutter [off-list ref] writes:

Hi Eric,

Thanks for your quick and insightful reply rightfully pointing out the
lack of rationale behind this change. So let me try to catch up:

Grr.  I did not get what you are trying to accomplish the first time I
skimmed this and rereading it all again closely I still don't get what
you are trying to acomplish.

What real world scenario do you have that approximates 100 mount
namespaces all sharing with each other with 1000 network namespaces
in that shared world?

I am inclined to suspect you are setting up containers that don't
contain and those 100 mount namespaces that share with each other
are your real concern.  But I don't know.

On Tue, Jul 05, 2016 at 09:44:00AM -0500, Eric W. Biederman wrote:

quoted

Phil Sutter [off-list ref] writes:

quoted

Stress-testing OpenStack Neutron revealed poor performance of 'ip netns'
when dealing with a high amount of namespaces. The cause of this lies in
the combination of how iproute2 mounts NETNS_RUN_DIR and the netns files
therein and the fact that systemd makes all mount points of the system
shared.

So please tell me.  Given that it was clearly a deliberate choice in the
code to make these directories shared, and that this is not a result
of a systemd making all directories shared by default.  Why is it
better to these directories non-shared?

NETNS_RUN_DIR itself is kept shared as it was intended by you (I hope).
The only difference is that we should avoid it being in the same group
as the parent mount point. Otherwise, all netns mount points will occur
twice.

How do they occur twice?  Are you dealing with a system that bind mounts
/run and /var/run?  The netns mount points occurring twice sounds
correct in that scenario.  Replacing a bind mount with a symlink would
be a more appropriate fix if you are concerned with the mount overhead.

Regarding the shared state of the netns mount points, I have actually no
idea what's the benefit, as there won't be any child mount points and
therefore no propagation should occur. Or am I missing something?

I think the second patch is probably ok.  I get turned around with the
finer points of mount propagation somedays as it is the parent mount
whose attributes matter when it comes to propagating the children.  

Still if the change semantically does not matter we have a missing
optimization in the kernel, and I would much rather implement that
optmization in the kernel than in every application that might possibly
hit it.  Especially given that the default on systemd systems is
"mount --make-rshared /"

quoted

This may be the appropriate change but saying you stress testing things
and have a problem but do not describe how large a scale you had a
problem, or anything else to make your problem reproducible by anyone
else makes it difficult to consider the merits of this change.

Sometimes things are a good default policy but have imperfect scaling on
extreme workloads.

My experience with the current situtation with ip netns is that it
prevents a whole lot of confusion by making the network namespace names
visible whichever mount namespace your processes are running in.

The only functional difference I noticed was the no longer twice
appearing netns mount points. They are still visible in all namespaces
though, just as before.

But you are fighting the how the rest of the system is configured at
that point and that concerns me.  iproute is not the place to
reconfigure the system.

Here's the script I wrote to benchmark 'ip netns':

| #!/bin/bash
| 
| IP=${IP:-/usr/sbin/ip}
| echo "using ip at $IP"
| 
| # make sure we start at a clean state
| for netns in $(ls /run/netns/* 2>/dev/null); do
|         $IP netns del ${netns##*/}
| done
| umount /run/netns
| 
| echo "creating 100 mount ns"
| touch /tmp/stay_alive
| for ((i = 0; i < 100; i++)); do
|         unshare -m --propagation unchanged bash -c \
| 		"while [[ -e /tmp/stay_alive ]]; do sleep 1; done" &
| done
| # give a little time for unshare to complete
| sleep 3
| 
| nscount=1000
| 
| echo -en "\ncreating $nscount netns"
| time (for ((i = 0; i < $nscount; i++)); do $IP netns add test$i; done)
| 
| echo -en "\ndeleting $nscount netns"
| time (for ((i = 0; i < $nscount; i++)); do $IP netns del test$i; done)
| 
| echo "removing mount ns again"
| rm /tmp/stay_alive
| wait

So basically it creates 100 idle mount namespaces, then times
adding/removing 1000 network namespaces. I called it three times:
without any patch, with just patch 1 and with both patches applied. Here
are the results:

| # IP=/tmp/base/ip /vmshare/reproducer/ip_netns_bench.sh
| using ip at /tmp/base/ip
| creating 100 mount ns
| 
| creating 1000 netns
| real	0m8.110s
| user	0m1.143s
| sys	0m6.235s
| 
| deleting 1000 netns
| real	0m15.347s
| user	0m0.957s
| sys	0m11.359s
| removing mount ns again

| # IP=/tmp/p1/ip /vmshare/reproducer/ip_netns_bench.sh
| using ip at /tmp/p1/ip
| creating 100 mount ns
| 
| creating 1000 netns
| real	0m7.956s
| user	0m0.987s
| sys	0m4.896s
| 
| deleting 1000 netns
| real	0m7.407s
| user	0m1.165s
| sys	0m3.418s
| removing mount ns again

| # IP=/tmp/p2/ip /vmshare/reproducer/ip_netns_bench.sh
| using ip at /tmp/p2/ip
| creating 100 mount ns
| 
| creating 1000 netns
| real	0m7.843s
| user	0m0.977s
| sys	0m4.915s
| 
| deleting 1000 netns
| real	0m6.407s
| user	0m1.006s
| sys	0m3.057s
| removing mount ns again

As you can see, the biggest improvement comes during deletion and from
patch 1. Though the second patch lowers the total time to delete the
namespaces by another second, which is still relatively much in
comparison to the low total time.

Which all seems to be about making /run/netns and /var/run/netns not
shared with each other which appears to be semantically wrong.

Eric