Theoretically HPC already has Ifniband; yet like most things HPC, it is fast because it is a trivial, unstable crap with every possible corner cut off -- it is basically just DMA over cable, app says "I want this buffer piped to that computer's RAM at this address", and IB hardware just do this async, bypassing network stack of the kernel. As one can imagine, this is super rigid and anti-resilient... And recently also nVidia-owned.
Though, IB was designed in times when CPU throughput was an inescapable bottleneck, I can imagine that one can now offload the standard Ethernet/IP stack to smarter NICs, abstract DMA with io-uring and get IB speeds and Ethernet's versatility in one thing, so maybe this is what they are after.
Though, IB was designed in times when CPU throughput was an inescapable bottleneck, I can imagine that one can now offload the standard Ethernet/IP stack to smarter NICs, abstract DMA with io-uring and get IB speeds and Ethernet's versatility in one thing, so maybe this is what they are after.
Comment