Re: [PATCH net-next 2/2] net: exit busy loop when another process is runnable

["Michael S. Tsirkin" <mst@redhat.com>]

On Wed, Sep 03, 2014 at 09:49:10AM +0300, Eliezer Tamir wrote:
> On 02/09/2014 11:31, Michael S. Tsirkin wrote:
> > On Tue, Sep 02, 2014 at 09:15:18AM +0300, Eliezer Tamir wrote:
> >> On 02/09/2014 06:29, Jason Wang wrote:
> >>> On 09/01/2014 02:39 PM, Eliezer Tamir wrote:
> >>>> On 29/08/2014 06:08, Jason Wang wrote:
> >>>>>> Yes, but rx busy polling only works in process context and does not
> >>>>>> disable bh, so it may be not an issue.
> >>>> sk_busy_loop() uses rcu_read_lock_bh(), so it does run with bh disabled.
> >>>
> >>> True, so we need probably also exit the loop when there are pending bhs.
> >>
> >> I'm not so sure, in the typical busy poll scenario, the incoming
> >> traffic is the most time-critical thing in the system.
> >> It's so important that you are willing to trade lots of CPU power
> >> for better latency. The user has decided that he wants to dedicate
> >> this CPU mostly for that. This is not something that plays nice with
> >> other apps, but this is what the user wants.
> > 
> > I think most applications wouldn't interpret this flag as "burn up CPU I don't
> > care what is the result", what apps want is more of "maximise throughput
> > and minimise latency even if throughput/CPU ratio goes down".
> > Jason posted benchmarks that show throughput going up because other
> > processes get more of a chance to run, so this seems consistent
> > with that goal.
> 
> Busy polling is not a general purpose feature, it's not something you
> can casually turn on and will "just work". Most applications should not
> be using busy polling. Currently it is used by multiserver applications
> that you spend days tuning to specific platforms.
> 
> What the user wants is to lower both avg and maximum latencies, at the
> expense of everything else including power efficiency and sometimes
> even throughput. The only exception is making the system crash ;)
> 
> While letting other things take precedence over busy polling might not
> hurt the avg latency much, it will kill your maximum latency.
> 
> -Eliezer

If scheduler happens to run both server and client on the
same CPU, polling will hurt maximum latency even more.
So I guess different users want different things.

How about applications giving us a hint what they prefer?
For example, a new flag that says "I don't have anything useful to do so
let's do busy polling but my server is on the local system, so please
only poll if CPU is otherwise idle".


-- 
MST