Re: [dpdk-dev] [PATCH v5 6/8] net/ice: support Rx AVX2 vector

["Lu, Wenzhuo" <wenzhuo.lu@intel.com>]

Hi Maxime,


> -----Original Message-----
> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> Sent: Wednesday, March 27, 2019 3:50 PM
> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev@dpdk.org
> Subject: Re: [dpdk-dev] [PATCH v5 6/8] net/ice: support Rx AVX2 vector
> 
> 
> 
> On 3/27/19 1:56 AM, Lu, Wenzhuo wrote:
> > Hi Maxime,
> >
> >> -----Original Message-----
> >> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> >> Sent: Tuesday, March 26, 2019 5:29 PM
> >> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev@dpdk.org
> >> Subject: Re: [dpdk-dev] [PATCH v5 6/8] net/ice: support Rx AVX2
> >> vector
> >>
> >> Hi,
> >>
> >> On 3/26/19 2:00 AM, Lu, Wenzhuo wrote:
> >>> Hi Maxime,
> >>>
> >>>> -----Original Message-----
> >>>> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> >>>> Sent: Monday, March 25, 2019 4:26 PM
> >>>> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev@dpdk.org
> >>>> Subject: Re: [dpdk-dev] [PATCH v5 6/8] net/ice: support Rx AVX2
> >>>> vector
> >>>>
> >>>> Hi,
> >>>>
> >>>> On 3/25/19 3:22 AM, Lu, Wenzhuo wrote:
> >>>>> Hi Maxime,
> >>>>>
> >>>>>
> >>>>>> -----Original Message-----
> >>>>>> From: Maxime Coquelin [mailto:maxime.coquelin@redhat.com]
> >>>>>> Sent: Friday, March 22, 2019 6:12 PM
> >>>>>> To: Lu, Wenzhuo <wenzhuo.lu@intel.com>; dev@dpdk.org
> >>>>>> Subject: Re: [dpdk-dev] [PATCH v5 6/8] net/ice: support Rx AVX2
> >>>>>> vector
> >>>>>
> >>>>>
> >>>>>>> +#ifndef RTE_LIBRTE_ICE_16BYTE_RX_DESC
> >>>>>>
> >>>>>> I see same is done for other Intel NICs, but I wonder what would
> >>>>>> be the performance cost of making it dynamic, if any cost?
> >>>>> Currently we don't have a good idea to make it dynamic. If we use
> >>>>> pointer
> >>>> to point to different functions for 16 byte and 32 byte, there's
> >>>> too much duplicate code to make it hard to maintain. If we use the
> >>>> same function, and check the configure in it. It impacts the
> performance.
> >>>>
> >>>> Have you done some measurements, what would be the performance
> >>>> impact?
> >>> I mean if we check the configuration is 16 byte or 32 byte, this
> >>> check will
> >> consume extra CPU cycles.
> >>> That why I think the better way is to have different paths for 16
> >>> byte and
> >> 32 byte. We should choose the appropriate path at the beginning.
> >>>
> >>>>
> >>>>> As HW does not support to change the configuration dynamically.
> >>>>> The
> >>>> device must be stopped and restarted if the configuration is changed.
> >>>> It's not very helpful to make it a dynamic configuration. We assume
> >>>> that the users can make their choice at the beginning and will not
> >>>> change
> >> it.
> >>>>
> >>>> The problem is that the user has to recompile to switch between the
> >>>> two configurations. And it may not be an option for the user if he
> >>>> uses dpdk packaged by a distribution, for example.
> >>>>
> >>>> Maybe I was not clear, but I don't mean to be able to switch mode
> >>>> while the port is started. I think it would be better to make it
> >>>> possible to switch mode at application startup time.
> >>> Yes, I understand the problem is the recompiling. But we think the
> >>> users
> >> will not change it after they made decision. That's why's acceptable
> >> in previous drivers.
> >>
> >> The problem is that the user may not be able to change it, if he does
> >> not get DPDK from source but from a distribution like Debian, Ubuntu or
> Red Hat.
> >>
> >> In this case, it means the user has no choice than sticking to 32
> >> bytes descriptors.
> > Normally using 32 bytes is the default behavior and it's good to do that.
> > But I have to say I don't quite understand the scenario. DPDK is open
> source, whatever OS that users are using, nothing prevents them going to
> dpdk website to get the code and customize it.
> 
> The user may prefer to use the distribution package for several reasons.
> Like not loosing the support he pays to the distributor by recompiling the
> package, or also not benefiting from the validation done by the distributor
> on the pre-built package.
Thanks for sharing the info about the deployment. Good to know that. How to say I'm always looking at the problem from the developer's point of view.

> 
> For example, would it make sense to fix the queue size at build time instead
> of using the --txd/--rxd run-time paramaters to save a few cycles here and
> there? I think not.
Agree, we have to balance them if we cannot make it perfect. We'll think about how to optimize it.