On Thu, Oct 13, 2022 at 3:23 PM Mattias Rönnblom
<mattias.ronnb...@ericsson.com> wrote:
>
> On 2022-10-13 03:32, Naga Harish K, S V wrote:
> >
> >
> >> -----Original Message-----
> >> From: Jayatheerthan, Jay <jay.jayatheert...@intel.com>
> >> Sent: Tuesday, October 11, 2022 12:40 PM
> >> To: mattias.ronnblom <mattias.ronnb...@ericsson.com>; Carrillo, Erik G
> >> <erik.g.carri...@intel.com>; Gujjar, Abhinandan S
> >> <abhinandan.guj...@intel.com>; Jerin Jacob <jer...@marvell.com>; Naga
> >> Harish K, S V <s.v.naga.haris...@intel.com>
> >> Cc: dev@dpdk.org; Van Haaren, Harry <harry.van.haa...@intel.com>;
> >> hof...@lysator.liu.se; mattias.ronnblom <mattias.ronnb...@ericsson.com>
> >> Subject: RE: [PATCH 2/4] eventdev: have ethernet Rx adapter appropriately
> >> report idle
> >>
> >> @Harish, Could you review the patch ?
> >>
> >> -Jay
> >>
> >>>
> >>> done:
> >>> - rx_adapter->stats.rx_intr_packets += nb_rx;
> >>> + if (nb_rx > 0) {
> >
> > How are the performance numbers before and after this patch?
> > Trying to understand the performance impact, as new condition is added to
> > the service function Datapath.
> >
> I haven't tested the RX and TX adapters separately, but if you run them
> on the same core, I get the following result:
>
> Without patches, with stats disabled: 16,0 Mpps
> Without patches, with stats enabled: 16,1 Mpps
> With patches, with stats disabled: 16,1 Mpps
> With patches, with stats enabled: 16,2 Mpps
>
> So these patches, with this particular hardware, compiler, and test
> application, adding a tiny bit of additional logic actually make the
> RX+TX adapter perform better. This is contrary to what you might expect,
> and I'm sure YMMV.
>
> Enabling service core statistics (which boils down to a 2x rdtsc and
> some cheap arithmetic in rte_service.c) actually make the RX+TX adapter
> core perform better, both before and after this patchset. Also contrary
> to what you might expect.
>
> The results are consistent across multiple runs.
>
> GCC 11.2.0 and AMD Zen 3 @ 3,7 GHz. Event device is DSW and I/O is the
> ring Ethdev.
@Naga Harish K, S V @Jayatheerthan, Jay @Gujjar, Abhinandan S @Erik
Gabriel Carrillo
Planning to take this series for rc2. If there are no other comments,
I will merge the series then.
