On Fri, Jan 04, 2019 at 07:58:09PM +0000, Hefty, Sean wrote: > > > I would agree if we can figure out a way to avoid the libibverbs API > > > overhead for fast path operations. > > > > This has been available for years now, the 'dv' interface does exactly > > this sort of bypass. The Mellanox drivers in ucx, dpdk, vma, etc all > > use the dv path for datapath operations. > > Something more generic that would avoid per device interfaces would be ideal. We have that, post_send/post_recv. If that is too slow then you have to use device specific DV, which has no limitations on dataplane performance. > > I think that is a consensus that EFA should build its libfabric > > provider on libibverbs & 'efadv'.. > > The user space implementation should be an orthogonal discussion > from the kernel. Why? We have to know what to review and kernel code is not going to be accepted without a user space implementation. Jason
