[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Message-ID: <6A6F17150A046D2D2670724D@nimrod.local>
Date: Sun, 08 May 2011 14:06:07 +0100
From: Alex Bligh <alex@...x.org.uk>
To: paulmck@...ux.vnet.ibm.com
cc: Eric Dumazet <eric.dumazet@...il.com>, netdev@...r.kernel.org,
Alex Bligh <alex@...x.org.uk>
Subject: Re: Scalability of interface creation and deletion
--On 8 May 2011 05:44:28 -0700 "Paul E. McKenney"
<paulmck@...ux.vnet.ibm.com> wrote:
> OK, let's break it out...
>
> 4.022 seconds for 100 interfaces means about 40 milliseconds per
> interface.
It's 200 interfaces as they are veth pairs, so 20ms.
> My guess is that you have CONFIG_NO_HZ=y, which means that RCU needs to
> figure out that various CPUs are in dyntick-idle state, which is a minimum
> of 6 jiffies.
That sounds like the main cause.
> It could be longer if a given CPU happens to be in IRQ
> when RCU checks, so call it 9 jiffies.
That sounds unlikely to happen much of the time. Looking at /proc/interrupts
not much is going on.
> If you are doing the interfaces
> synchronously, you will likely have to wait for a prior grace period (due
> to background activity).
Not sure I understand that (there's nothing else going on), but...
> So I can easily imagine 18 milliseconds for
> HZ=1000. 40 milliseconds sounds a bit high, but perhaps not impossible.
Add to that the fact that there are 2 x synchronize_sched() and one
rcu_barrier() per interface removal. If I ignore your IRQ idea, that's
6 x 3 = 18ms per interface removal at CONFIG_HZ=10, 180ms at
CONFIG_HZ=100.
--
Alex Bligh
--
To unsubscribe from this list: send the line "unsubscribe netdev" in
the body of a message to majordomo@...r.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Powered by blists - more mailing lists