Re: [PATCH net-next v4] rps: selective flow shedding during softnet overflow

[Eric Dumazet <eric.dumazet@gmail.com>]

On Tue, 2013-04-23 at 18:41 -0400, David Miller wrote:
> From: Willem de Bruijn <willemb@google.com>
> Date: Tue, 23 Apr 2013 16:31:34 -0400
> 
> > - disable the kconfig option by default, as it is workload specific.
> 
> If this logic only kicks in during overload, I see no reason why we
> shouldn't have this protection enabled unconditionally, all the time.
> 
> Does it hurt performance under normal workloads?
> 
> If the new logic is controlled by overload thresholds then it should
> have no impact whatsoever in non-overload scenerios.

Exact.

I guess we should at least not include this code on !SMP builds

This adds some code, but as long as you don't write into
flow_limit_cpu_bitmap/flow_limit_table_len we wont allocate the memory,
thus this is disabled.

Then, if enabled, it only kicks if the number of queued packets is above
half the max backlog (netdev_max_backlog >> 1)

So on non overload scenario, cost is 0.

On overload scenario, the loaded cpu keeps a 512 bytes array hot in his
cache. (fl->history[])