Re: [dpdk-dev] Limiting packet buffers under 4GB

[Jerin Jacob <jerinjacobk@gmail.com>]

On Wed, 6 Nov, 2019, 5:46 pm Burakov, Anatoly, <anatoly.burakov@intel.com>
wrote:

> On 04-Nov-19 12:59 PM, Jerin Jacob Kollanukkaran wrote:
> > Hi Anatoly and All,
> >
> > Just wondering what would the side effect of lowering a  _bit_ of static
> uint64_t baseaddr = 0x100000000 in
> > lib/librte_eal/common/eal_common_memory.c for 64bit systems.
> >
> > Use case:
> > If we _reserve_ VA address which less than 2^32 ONLY for packet
> buffers(mbuf), The use cases like
> > Pipeline, where need to transfer packets from one core to another cores
> can use ring element
> > size of 4B(32bit) which will reduce the a lot of read and write  to
> enable better
> > performance.
> >
> > i.e Since upper 32bits will be zero, it is matter of typecasting of item
> to read and write from/to ring.
> > Essentially memcpy overhead for moving pointers over the ring will be
> half.
> >
> > Is baseaddr set to 2^32 to make sure that secondary process will have
> more _chance_ of getting
> > the baseaddr in order for DPDK to work?
> >
> > Thoughts on above? On general to reduce the mbuf pointer storage
> requirement for ring?
> >
>
> You can already try that with --base-virtaddr option if you have such
> specific addressing requirements. That said, the address is pretty much
> arbitrary, so i'm not against lowering it in principle, it just feels
> like a workaround for something that's very specific to the workloads
> you are targetting.
>

Thanks for the feedback. I will find the performance delta first to decide
any such option makes sense.



> --
> Thanks,
> Anatoly
>