> On Jul 30, 2019, at 10:54 AM, Stephen Hemminger <step...@networkplumber.org>
> wrote:
>
> On Tue, 30 Jul 2019 16:33:55 +0100
> Bruce Richardson <bruce.richard...@intel.com> wrote:
>
>> On Tue, Jul 30, 2019 at 08:25:34AM -0700, Stephen Hemminger wrote:
>>> On Tue, 30 Jul 2019 14:49:49 +0200
>>> Marcin Zapolski <marcinx.a.zapol...@intel.com> wrote:
>>>
>>>> Make rte_eth_rx_burst, rte_eth_tx_burst and other static inline ethdev
>>>> functions not inline. They are referencing DPDK internal structures and
>>>> inlining forces those structures to be exposed to user applications.
>>>>
>>>> In internal testing with i40e NICs a performance drop of about 2% was
>>>> observed with testpmd.
>>>>
>>>> Signed-off-by: Marcin Zapolski <marcinx.a.zapol...@intel.com>
>>>
>>> Sorry 2% matters.
>>
>> Note that this is with testpmd. Are there many apps out there where a 2%
>> drop in IO cost would be noticable?
>
> Why not find a way to get the 2% back elsewhere? Maybe analyzing the
> code/cache
> in more detail. Perhaps some prefetching could help, or getting rid of
> indirect calls elsewhere in the code. At the extreme, maybe implementing
> something like the kernel static branches (self-modifying code) would
> get a lot back.
+1, I discussed something very similar internally or at least let's not reduce
DPDK performance by 2% and find a different way.
Regards,
Keith
