New optimized soreceive_stream() for TCP sockets, proof of concept

Andrew Gallatin gallatin at cs.duke.edu
Mon Mar 5 19:58:45 UTC 2007


One last note..  It looks like SCHED_4BSD does a decent job (on my
setup) w/o CPU binding, but SCHED_ULE requires CPU binding to get good
performance.  W/o CPU binding, the best bandwidth I see using
SCHED_ULE is around 5.3Gb/s with one CPU mostly idle..  With CPU
binding, it is roughly 9Gb/s.  

Drew


More information about the freebsd-net mailing list