3 ms·

> You can now get a network pipe in to a system that a rather beefy multi-core CPU using a user-space TCP stack can barely keep up with. I'd love to learn more

• jerf 9y ago
  Google around for the Intel DPDK, and you can find things like this: http://www.cs.cornell.edu/courses/cs5413/2014fa/projects/group_of_dsd96_vs444/final_doc.pdf http://www.cs.cornell.edu/courses/cs5413/2014fa/projects/gro... You can also do a simple math analysis to see it. If you have an incoming 10Gbps connection, a single core machine has approx. 1/3rd of a cycle per bit to do everything it's going to do with that packet. Even going to a 128 core machine and assuming perfect parallelism with some sort of magical packet muxer gives you a whopping 43-ish cycles per bit. I've never worked on this myself, but I saw a team in my company working with it and were pretty pleased to be able to push ~2Gbps through their 10Gbps network connection with a pretty beefy machine, and just about all they were doing was relatively simple load balancing.
• mtanski 9y ago
  LWN has had quite a few past articles that went into details of issues with processing network packets at line speeds on modern OSes and machines. https://lwn.net/Articles/629155/ https://lwn.net/Articles/629155/ https://lwn.net/Articles/713918/ https://lwn.net/Articles/713918/ https://lwn.net/Articles/719850/ https://lwn.net/Articles/719850/ The block layer has gone through somewhat similar issues as some storage devices started approaching RAM speeds.