Re: [dpdk-dev] [PATCH RFC 0/3] DPDK ethdev callback support

[Bruce Richardson <bruce.richardson@intel.com>]

On Mon, Dec 22, 2014 at 05:33:07PM +0000, Bruce Richardson wrote:
> On Mon, Dec 22, 2014 at 06:02:53PM +0100, Thomas Monjalon wrote:
> > Hi Bruce,
> > 
> > Callbacks, as hooks for applications, give more flexibility and are
> > generally a good idea.
> > In DPDK the main issue will be to avoid performance degradation.
> > I see you use "unlikely" for callback branching.
> > Could we reduce more the impact of this test by removing the queue array,
> > i.e. having port-wide callbacks instead of per-queue callbacks?
> 
> I can give that a try, but I don't see it making much difference if any. The
> main thing to avoid with branching is branch mis-prediction, which should not
> be a problem here, as the user is not going to be adding or removing callbacks
> between each RX and TX call, making the branches highly predictable - i.e. always
> go the same way.
> The reason for using per-queue callbacks is that I think we can do more with
> it that way. For instance, if we want to do some additional processing or
> calculations on only IP traffic, then we can use hardware offloads on most
> NICs to steer the IP traffic to a separate queue and only apply the callbacks
> to that queue. If the performance is the same, I think we should therefore keep
> the per-queue version.
> 

I tried just using a single callback pointer per port, instead of per-queue. 
Performance looked the same for the no-callback case, the case of an empty
callback function, and the test application with callbacks for latency
computation.

> > 
> > 2014-12-22 16:47, Bruce Richardson:
> > > Future extensions: in future the ethdev library can be extended to provide
> > > a standard set of callbacks for use by drivers. 
> > 
> > Having callbacks for drivers seems strange to me.
> > If drivers need to accomplish some tasks, they do it by implementing an
> > ethdev service. New services are declared for new needs.
> > Callbacks are the reverse logic. Why should it be needed?
> 
> Typo, I meant for applications! Drivers don't need them indeed.
> 
> > 
> > > For now this patch set is RFC and still needs additional work for creating
> > > a remove function for callbacks and to add in additional testing code.
> > > Since this adds in new code into the critical data path, I have run some
> > > performance tests using testpmd with the ixgbe vector drivers (i.e. the
> > > fastest, fast-path we have :-) ). Performance drops due to this patch
> > > seems minimal to non-existant, rough tests on my system indicate a drop
> > > of perhaps 1%.
> > > 
> > > All feedback welcome.
> > 
> > It would be good to have more performance tests with different configurations.
> 
> Sure, if you have ideals for specific tests you'd like to see I'll try and
> get some numbers. What I did look as was the performance impact for this patch
> without actually putting in place any callbacks, and the worst-case here is
> hardly noticable. For an empty callback, i.e. the pure callback overhead, the
> performance should still be in low single-digit percentages, but I'll test to 
> confirm that. For other slower RX and TX paths, e.g. those using scattered
> packets, or with TX offloads, the performance impact will be even less.
>

I modified the sample app in patch 3 of this set to have empty callbacks on
both RX and TX, and the performance impact vs the non-callback case was again
about 1% only. [Again using the vector RX/TX fastpath functions, with 4 ports
handled by a single lcore].

Regards,
/Bruce