User-space TCP stack: lavrd/tunnel
NetworkingComments
The hybrid threading engine is a thoughtful addition. I am curious if the synchronization overhead between these threading models might offset some of the latency gains from skipping the kernel.
cache misses from thread migration usually cost more than the syscall itself.
User-space stacks are interesting in theory, but they often struggle when you have to deal with shared network interfaces or legacy hardware in a real rack. Performance on a clean VM is rarely the same as performance on a production server with a dozen other services fighting for the NIC.
If this is intended for environments with specialized hardware like DPDK enabled NICs, would the kernel's networking overhead still be the primary bottleneck? Perhaps the gain is less about raw throughput and more about the deterministic timing of packet processing.