[dpdk-dev] [PATCH] eal: use get_mempolicy(2) to find numa socket on Linux

Thomas Monjalon thomas at monjalon.net
Mon Jul 3 16:20:20 CEST 2017


03/07/2017 16:11, Sergio Gonzalez Monroy:
> On 01/07/2017 15:18, Thomas Monjalon wrote:
> > 21/06/2017 17:49, Sergio Gonzalez Monroy:
> >> I think following the discussion of libnuma dependency in another thread
> >> [1], you would need to implement a similar approach and keep the old
> >> method while providing this alternative if libnuma is present.
> > Considering that libnuma becomes mandatory to build DPDK on NUMA-capable
> > systems, we can assume there is only one CPU socket if libnuma is
> > unavailable.
> >
> >> As Ilya mentions in the thread, this is usually the job of tools such as
> >> autotools, cmake or meson but given that we do not have such tools in
> >> DPDK yet, we rely in a build time config option for libnuma.
> >>
> >> Given that we already have a libnuma config option for VHOST, we might
> >> be adding anew one for hugepage balancing, I think it would make sense
> >> to just have single CONFIG_RTE_LIBNUMA option instead.
> >>
> >> Thoughts?
> > I am not sure about merging every NUMA options into one.
> > It may be interesting to track different NUMA features requiring libnuma.
> 
> Do you see the case where a system with libnuma we want to use the 
> support for hugapage allocation but not for vhost?
> In my opinion this looks like a manual check for libnuma where with a 
> different build system would have been automated.

I think it is not related to the build system.
We could have an automatic check for libnuma which enables several options.
I'm just advocating to keep the different uses of libnuma clearly
advertised in order to simply find what are the features enabled
with this dependency.




More information about the dev mailing list