Re: Scalability of interface creation and deletion

--On 7 May 2011 20:32:54 +0200 Eric Dumazet [off-list ref] wrote:

Well, there is also one rcu_barrier() call that is expensive.
(It was changed from one synchronize_rcu() to one rcu_barrier() lately
in commit ef885afb , in 2.6.36 kernel)

I think you are saying it may be waiting in rcu_barrier(). I'll
instrument that later plus synchronize_sched().

http://git2.kernel.org/?p=linux/kernel/git/torvalds/linux-2.6.git;a=commi
tdiff;h=ef885afbf8a37689afc1d9d545e2f3e7a8276c17

OK, so in head, which I am using, rollback_registered_many which
previously had 2 calls to synchronize_net(), now has one, followed
by a call to rc_barrier() at the bottom.

netdev_wait_allrefs() waits that all references to a device vanishes.

It currently uses a _very_ pessimistic 250 ms delay between each probe.
Some users reported that no more than 4 devices can be dismantled per
second, this is a pretty serious problem for some setups.

Right, that's what I patched before (see patch attached to
message from earlier today) to do an exponential backoff (see
previous entry), i.e. do a 5ms sleep, then a 10ms, then a 20ms, but
never more than 250ms. It made no difference.

time to remove 50 ipip tunnels on a UP machine :

before patch : real 11.910s
after patch : real 1.250s

Sadly I don't see that improvement!

-- 
Alex Bligh