Optimizing the kernel to saturate a 100Gbps link (2017)

Novel concept: Include the described link so every curious soul doesn't have to dig.

https://news.ycombinator.com/item?id=28584738

Just sayin', clickable links are pretty nice.

one of the things I worry about is that with mellanox now being owned by nvidia, they're going to be much more hostile in the future to having real open source full capability drivers for the linux and freebsd kernels.

the drivers you can download from the mellanox website to build them as DKMS modules already have absurd licenses.

I’ve come across those but - perhaps because they’re just slides - they really don’t go into the same depth the 100Gbps article does. They’re amazing resources but a little more high level and feature very little kernel hacking.

Every time I'm on a box with 1+ GPUs and 10-100GB/s of PCIe, I keep thinking, "we'll give you more after you've used what we've got". New tech like GPU Direct is amazing, and while the tools are surprisingly easy, the real work is getting sw archs to have individual parts think in 100GB/s. Exciting times!

In the context of Netflix, Intel NIC doesn't seems to offers kTLS acceleration?

Userspace isn't magic. (Unfortunately the "userspace networking" people have done a poor job of explaining it.) When you're serving static files like Netflix, you want the storage drivers, filesystem, cache, network stack, and NIC drivers to be in the same address space. Whether that address space is userspace or kernel is not that important.