[dpdk-dev] [PATCH v2 1/2] ethdev: add buffered tx api

[Thomas Monjalon]

2016-03-09 16:17, Ananyev, Konstantin:
> From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> > 2016-03-09 15:42, Ananyev, Konstantin:
> > > From: Thomas Monjalon [mailto:thomas.monjalon at 6wind.com]
> > > > 2016-03-09 15:23, Ananyev, Konstantin:
> > > > > >
> > > > > > 2016-03-09 13:36, Ananyev, Konstantin:
> > > > > > > > > +   if (to_send == 0)
> > > > > > > > > +           return 0;
> > > > > > > >
> > > > > > > > Why this check is done in the lib?
> > > > > > > > What is the performance gain if we are idle?
> > > > > > > > It can be done outside if needed.
> > > > > > >
> > > > > > > Yes, that could be done outside, but if user has to do it anyway,
> > > > > > > why not to put it inside?
> > > > > > > I don't expect any performance gain/loss because of that -
> > > > > > > just seems a bit more convenient to the user.
> > > > > >
> > > > > > It is handling an idle case so there is no gain obviously.
> > > > > > But the condition branching is surely a loss.
> > > > >
> > > > > I suppose that condition should always be checked:
> > > > > either in user code prior to function call or inside the
> > > > > function call itself.
> > > > > So don't expect any difference in performance here...
> > > > > Do you have any particular example when you think it would?
> > > > > Or are you talking about rte_eth_tx_buffer() calling
> > > > > rte_eth_tx_buffer_flush() internally?
> > > > > For that one - both are flush is 'static inline' , so I expect
> > > > > compiler be smart enough to remove this redundant check.
> > > > >
> > > > > > So why the user would you like to do this check?
> > > > > Just for user convenience - to save him doing that manually.
> > > >
> > > > Probably I've missed something. If we remove this check, the function
> > > > will do nothing, right? How is it changing the behaviour?
> > >
> > > If we'll remove that check, then
> > > rte_eth_tx_burst(...,nb_pkts=0)->(*dev->tx_pkt_burst)(...,nb_pkts=0)
> > > will be called.
> > > So in that case it might be even slower, as we'll have to do a proper call.
> > 
> > If there is no packet, we have time to do a useless call.
> 
> One lcore can do TX for several queues/ports.
> Let say we have N queues to handle, but right now traffic is going only through
> one of them. 
> That means we'll have to do N-1 useless calls and reduce number of cycles
> available to send actual traffic.

OK, good justification, thanks.

> > > Of course user can avoid it by:
> > >
> > > If(tx_buffer->nb_pkts != 0)
> > > 	rte_eth_tx_buffer_flush(port, queue, tx_buffer);
> > >
> > > But as I said what for to force user to do that?
> > > Why not to  make this check inside the function?
> > 
> > Because it may be slower when there are some packets
> > and will "accelerate" only the no-packet case.
> > 
> > We do not progress in this discussion.
> > It is not a big deal, 
> 
> Exactly.
> 
> >just a non sense.
> 
> Look at what most of current DPDK examples do: they do check manually
> does nb_pkts==0 or not, if not call tx_burst().
> For me it makes sense to move that check into the library function -
> so each and every caller doesn't have to do it manually.
> 
> > So I agree to keep it if we change the website to announce that DPDK
> > accelerates the idle processing ;)
> 
> That's fine by me, but at first I suppose you'll have to provide some data
> showing that this approach slowdowns things, right? :)

You got me