Memory providers multiplexing (Was: [PATCH net-next v4 4/5] page_pool: remove PP_FLAG_PAGE_FRAG flag)

Jakub Kicinski kuba at kernel.org
Tue Jul 11 10:06:36 PDT 2023


On Tue, 11 Jul 2023 13:42:47 -0300 Jason Gunthorpe wrote:
> On Tue, Jul 11, 2023 at 09:00:47AM -0700, Jakub Kicinski wrote:
> > > So is Infiniband, Jakub has a unique definition of "proprietary".  
> > 
> > For IB AFAIU there's only one practically usable vendor, such an
> > impressive ecosystem!!  
> 
> IB has roce (RDMA Over Converged Ethernet, better thought of as IB
> over Ethernet) under it's standization umbrella and every major
> commodity NIC vendor (mellanox, broadcom, intel) now implements
> roce.

Now we're getting into our favorite argument and completely
sidetracking the conversation, aren't we? :) And as usual 
our ability to present facts is limited by various NDAs..

> We also have the roce support in the switch from all major
> switch vendors.

By which you mean all major switch vendors should support basic RoCE
requirements. But most vendors will try to put special features into
their switches trying to make the full NIC + switch solution as sticky
as possible.

> IB as a link layer "failed" with most implementors leaving the

Interesting.

> ecosystem broadly because Ethernet link layer always wins in
> networking - this is more of an economic outcome than a
> standardization outcome in my mind.
> 
> Due to roce, IB as a transport and software protocol has a solid
> multi-vendor ecosystem.
 
Last I checked every generation of HW from even a single vendor came out
with a new congestion control algorithm and add-ons. All closed source
and usually patented. That is the core of a transport protocol, the
intelligence, not the frame format (which may or may not itself deviate
thru various encaps and stolen bits).



More information about the Linux-mediatek mailing list