Re: [PATCH v8 bpf-next 00/14] mvneta: introduce XDP multi-buffer support

[Jesper Dangaard Brouer <brouer@redhat.com>]

On Tue, 20 Apr 2021 15:49:44 +0200
Magnus Karlsson <magnus.karlsson@gmail.com> wrote:

> On Mon, Apr 19, 2021 at 8:56 AM Lorenzo Bianconi
> <lorenzo.bianconi@redhat.com> wrote:
> >  
> > > On Sun, Apr 18, 2021 at 6:18 PM Jesper Dangaard Brouer
> > > <brouer@redhat.com> wrote:  
> > > >
> > > > On Fri, 16 Apr 2021 16:27:18 +0200
> > > > Magnus Karlsson <magnus.karlsson@gmail.com> wrote:
> > > >  
> > > > > On Thu, Apr 8, 2021 at 2:51 PM Lorenzo Bianconi <lorenzo@kernel.org> wrote:  
> > > > > >
> > > > > > This series introduce XDP multi-buffer support. The mvneta driver is
> > > > > > the first to support these new "non-linear" xdp_{buff,frame}. Reviewers
> > > > > > please focus on how these new types of xdp_{buff,frame} packets
> > > > > > traverse the different layers and the layout design. It is on purpose
> > > > > > that BPF-helpers are kept simple, as we don't want to expose the
> > > > > > internal layout to allow later changes.
> > > > > >
> > > > > > For now, to keep the design simple and to maintain performance, the XDP
> > > > > > BPF-prog (still) only have access to the first-buffer. It is left for
> > > > > > later (another patchset) to add payload access across multiple buffers.
> > > > > > This patchset should still allow for these future extensions. The goal
> > > > > > is to lift the XDP MTU restriction that comes with XDP, but maintain
> > > > > > same performance as before.  
> > > > [...]  
> > > > > >
> > > > > > [0] https://netdevconf.info/0x14/session.html?talk-the-path-to-tcp-4k-mtu-and-rx-zerocopy
> > > > > > [1] https://github.com/xdp-project/xdp-project/blob/master/areas/core/xdp-multi-buffer01-design.org
> > > > > > [2] https://netdevconf.info/0x14/session.html?tutorial-add-XDP-support-to-a-NIC-driver (XDPmulti-buffers section)  
> > > > >
> > > > > Took your patches for a test run with the AF_XDP sample xdpsock on an
> > > > > i40e card and the throughput degradation is between 2 to 6% depending
> > > > > on the setup and microbenchmark within xdpsock that is executed. And
> > > > > this is without sending any multi frame packets. Just single frame
> > > > > ones. Tirtha made changes to the i40e driver to support this new
> > > > > interface so that is being included in the measurements.  
> > > >
> > > > Could you please share Tirtha's i40e support patch with me?  
> > >
> > > We will post them on the list as an RFC. Tirtha also added AF_XDP
> > > multi-frame support on top of Lorenzo's patches so we will send that
> > > one out as well. Will also rerun my experiments, properly document
> > > them and send out just to be sure that I did not make any mistake.  
> >
> > ack, very cool, thx  
> 
> I have now run a new set of experiments on a Cascade Lake server at
> 2.1 GHz with turbo boost disabled. Two NICs: i40e and ice. The
> baseline is commit 5c507329000e ("libbpf: Clarify flags in ringbuf
> helpers") and Lorenzo's and Eelco's path set is their v8. First some
> runs with xdpsock (i.e. AF_XDP) in both 2-core mode (app on one core
> and the driver on another) and 1-core mode using busy_poll.
> 
> xdpsock rxdrop throughput change with the multi-buffer patches without
> any driver changes:
> 1-core i40e: -0.5 to 0%   2-cores i40e: -0.5%
> 1-core ice: -2%   2-cores ice: -1 to -0.5%
> 
> xdp_rxq_info -a XDP_DROP
> i40e: -4%   ice: +8%
> 
> xdp_rxq_info -a XDP_TX
> i40e: -10%   ice: +9%
> 
> The XDP results with xdp_rxq_info are just weird! I reran them three
> times, rebuilt and rebooted in between and I always get the same
> results. And I also checked that I am running on the correct NUMA node
> and so on. But I have a hard time believing them. Nearly +10% and -10%
> difference. Too much in my book. Jesper, could you please run the same
> and see what you get? 

We of-cause have to find the root-cause of the +-10%, but let me drill
into what the 10% represent time/cycle wise.  Using a percentage
difference is usually a really good idea as it implies a comparative
measure (something I always request people to do, as a single
performance number means nothing by itself).

For a zoom-in-benchmarks like these where the amount of code executed
is very small, the effect of removing or adding code can effect the
measurement a lot.

I can only do the tests for i40e, as I don't have ice hardware (but
Intel is working on fixing that ;-)).

 xdp_rxq_info -a XDP_DROP
  i40e: 33,417,775 pps

 CPU is 100% used, so we can calculate nanosec used per packet:
  29.92 nanosec (1/33417775*10^9)
  2.1 GHz CPU =  approx 63 CPU-cycles

 You lost -4% performance in this case.  This correspond to:
  -1.2 nanosec (29.92*0.04) slower
  (This could be cost of single func call overhead = 1.3 ns)
  
My measurement for XDP_TX:

 xdp_rxq_info -a XDP_TX
  28,278,722 pps
  35.36 ns (1/28278722*10^9)

 You lost -10% performance in this case:
  -3.54 nanosec (35.36*0.10) slower

In XDP context 3.54 nanosec is a lot, as you can see it is 10% in this
zoom-in benchmark.  We have to look at the details.

One detail/issue with i40e doing XDP_TX, is that I cannot verify that
packets are actually transmitted... not via exception tracepoint, not
via netstats, not via ethtool_stats.pl.  Maybe all the packets are
getting (silently) drop in my tests...!?!


> The xdpsock numbers are more in the ballpark of
> what I would expect.
>
> Tirtha and I found some optimizations in the i40e
> multi-frame/multi-buffer support that we have implemented. Will test
> those next, post the results and share the code.
> 
> > >
> > > Just note that I would really like for the multi-frame support to get
> > > in. I have lost count on how many people that have asked for it to be
> > > added to XDP and AF_XDP. So please check our implementation and
> > > improve it so we can get the overhead down to where we want it to be.  
> >
> > sure, I will do.
> >
> > Regards,
> > Lorenzo
> >  
> > >
> > > Thanks: Magnus
> > >  
> > > > I would like to reproduce these results in my testlab, in-order to
> > > > figure out where the throughput degradation comes from.
> > > >  
> > > > > What performance do you see with the mvneta card? How much are we
> > > > > willing to pay for this feature when it is not being used or can we in
> > > > > some way selectively turn it on only when needed?  
> > > >
> > > > Well, as Daniel says performance wise we require close to /zero/
> > > > additional overhead, especially as you state this happens when sending
> > > > a single frame, which is a base case that we must not slowdown.
> > > >
> > > > --
> > > > Best regards,
> > > >   Jesper Dangaard Brouer

-- 
Best regards,
  Jesper Dangaard Brouer
  MSc.CS, Principal Kernel Engineer at Red Hat
  LinkedIn: http://www.linkedin.com/in/brouer


Running XDP on dev:i40e2 (ifindex:6) action:XDP_DROP options:read
XDP stats       CPU     pps         issue-pps  
XDP-RX CPU      2       33,417,775  0          
XDP-RX CPU      total   33,417,775 

RXQ stats       RXQ:CPU pps         issue-pps  
rx_queue_index    2:2   33,417,775  0          
rx_queue_index    2:sum 33,417,775 


Running XDP on dev:i40e2 (ifindex:6) action:XDP_TX options:swapmac
XDP stats       CPU     pps         issue-pps  
XDP-RX CPU      2       28,278,722  0          
XDP-RX CPU      total   28,278,722 

RXQ stats       RXQ:CPU pps         issue-pps  
rx_queue_index    2:2   28,278,726  0          
rx_queue_index    2:sum 28,278,726