Re: Possible netns creation and execution performance/scalability regression since v3.8 due to rcu callbacks being offloaded to multiple cpus

[ebiederm@xmission.com (Eric W. Biederman)]

Rafael David Tinoco <rafael.tinoco@canonical.com> writes:

> Hello Eric,
>
> Coming back to this...
>
> On Jun 16, 2014, at 12:01 PM, Rafael Tinoco <rafael.tinoco@canonical.com> wrote:
>
>> ...
>> 
>> On Fri, Jun 13, 2014 at 9:02 PM, Eric W. Biederman
>> <ebiederm@xmission.com> wrote:
>>> Rafael Tinoco <rafael.tinoco@canonical.com> writes:
>>> 
>>>> Okay,
>>>> 
>>>> Tests with the same script were done.
>>>> I'm comparing : master + patch vs 3.15.0-rc5 (last sync'ed rcu commit)
>>>> and 3.9 last bisect good.
>>>> 
>>>> Same tests were made. I'm comparing the following versions:
>>>> 
>>>> 1) master + suggested patch
>>>> 2) 3.15.0-rc5 (last rcu commit in my clone)
>>>> 3) 3.9-rc2 (last bisect good)
>>> 
>>> I am having a hard time making sense of your numbers.
>>> 
>>> If I have read your email correctly my suggested patch caused:
>>> "ip netns add" numbers to improve
>>> 1x "ip netns exec" to improve some
>>> 2x "ip netns exec" to show no improvement
>>> "ip link add" to show no effect (after the 2x ip netns exec)
>> 
>> - "netns add" are as good as they were before this regression.
>> - "netns exec" are improved but still 50% of the last good bisect commit.
>> - "link add" didn't show difference.
>> 
>>> This is interesting in a lot of ways.
>>> - This seems to confirm that the only rcu usage in ip netns add
>>>  was switch_task_namespaces.  Which is convinient as that rules
>>>  out most of the network stack when looking for performance oddities.
>>> 
>>> - "ip netns exec" had an expected performance improvement
>>> - "ip netns exec" is still slow (so something odd is still going on)
>>> - "ip link add" appears immaterial to the performance problem.
>>> 
>>> It would be interesting to switch the "ip link add" and "ip netns exec"
>>> in your test case to confirm that there is nothing interesting/slow
>>> going on in "ip link add"
>> 
>> - will do that.
>
> IP link add seems ok.
>
>> 
>>> 
>>> Which leaves me with the question what ip "ip netns exec" remains
>>> that is using rcu and is slowing all of this down.
>> 
>> - will check this also.
>
> Based on my tests (and some other users that deployed this patch on a server farm)
> it looks like changing rcu_read_lock() to task_lock() did the trick. We are getting
> same (sometimes much better) results - comparing bisect good - for a big amount 
> of netns being created simultaneously. 
>
> Is it possible to make this change permanent in kernel tree ? 

Definitely.  I just need to finish getting my act together.

It sounded like you had seen other performance problems and I was
waiting on your futher testing so we could narrow down.

If you can't see other problems then I am happy to move forward with
this.

Thank you for testing and reporting this,

Eric