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

[Rafael Tinoco]

Eric,

I'll test the patch with the same testcase and let you all know.

Really appreciate everybody's efforts.

On Wed, Jun 11, 2014 at 5:55 PM, Eric W. Biederman
<ebiederm@xxxxxxxxxxxx> wrote:
> "Paul E. McKenney" <paulmck@xxxxxxxxxxxxxxxxxx> writes:
>
>> On Wed, Jun 11, 2014 at 01:27:07PM -0500, Dave Chiluk wrote:
>>> On 06/11/2014 11:18 AM, Paul E. McKenney wrote:
>>> > On Wed, Jun 11, 2014 at 10:46:00AM -0500, David Chiluk wrote:
>>> >> Now think about what happens when a gateway goes down, the namespaces
>>> >> need to be migrated, or a new machine needs to be brought up to replace
>>> >> it.  When we're talking about 3000 namespaces, the amount of time it
>>> >> takes simply to recreate the namespaces becomes very significant.
>>> >>
>>> >> The script is a stripped down example of what exactly is being done on
>>> >> the neutron gateway in order to create namespaces.
>>> >
>>> > Are the namespaces torn down and recreated one at a time, or is there some
>>> > syscall, ioctl(), or whatever that allows bulk tear down and recreating?
>>> >
>>> >                                                    Thanx, Paul
>>>
>>> In the normal running case, the namespaces are created one at a time, as
>>> new customers create a new set of VMs on the cloud.
>>>
>>> However, in the case of failover to a new neutron gateway the namespaces
>>> are created all at once using the ip command (more or less serially).
>>>
>>> As far as I know there is no syscall or ioctl that allows bulk tear down
>>> and recreation.  if such a beast exists that might be helpful.
>>
>> The solution might be to create such a beast.  I might be able to shave
>> a bit of time off of this benchmark, but at the cost of significant
>> increases in RCU's CPU consumption.  A bulk teardown/recreation API could
>> reduce the RCU grace-period overhead by several orders of magnitude by
>> having a single RCU grace period cover a few thousand changes.
>>
>> This is why other bulk-change syscalls exist.
>>
>> Just out of curiosity, what syscalls does the ip command use?
>
> You can look in iproute2 ip/ipnetns.c
>
> But rought ip netns add does:
>
> unshare(CLONE_NEWNET);
> mkdir /var/run/netns/<name>
> mount --bind /proc/self/ns/net /var/run/netns/<name>
>
> I don't know if there is any sensible way to batch that work.
>
> (The unshare gets you into copy_net_ns in net/core/net_namespace.c
>  and to find all of the code it can call you have to trace all
>  of the register_pernet_subsys and register_pernet_device calls).
>
> At least for creation I would like to see if we can make all of the
> rcu_callback synchronize_rcu calls go away.  That seems preferable
> to batching at creation time.
>
> Eric



-- 
-- 
Rafael David Tinoco
Software Sustaining Engineer @ Canonical
Canonical Technical Services Engineering Team
# Email: rafael.tinoco@xxxxxxxxxxxxx (GPG: 87683FC0)
# Phone: +55.11.9.6777.2727 (Americas/Sao_Paulo)
# LP: ~inaddy | IRC: tinoco | Skype: rafael.tinoco
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/