4 ms·

IO_uring Network Zero-Copy Receive Lands in Linux 6.15

• jauntywundrkind 2y ago
  > While this version only supports host memory as that was the initial target, other memory types are planned as well, with notably GPU memory coming next. Awhhh yeah. Return to PacketShader (2010)! Honestly not sure if there are use cases for GPU routing. More likely use case would probably be feeding LLMs data over the network with lower overhead. I'm not sure what's possible today with PCIe p2p dma, how much this io_uring capability could potentially improve those direct-to-GPU scenarios. https://docs.kernel.org/driver-api/pci/p2pdma.html https://docs.kernel.org/driver-api/pci/p2pdma.html
• pestatije 2y ago
  > I was able to saturate a 200G link with a single CPU core how can a socket receive be able to saturate a link?...you can only saturate a link when sending, not when receiving... what am i missing?
  • magicalhippo 2y ago
    There's nothing stating what they used on the sending side. So I read it as they were able to saturate a 200GBps link while not being receive limited despite restricting the receiver to a single core.