Semi-polling mode and net.inet.tcp.inflight_enable

Terry Lambert tlambert2 at mindspring.com
Thu Apr 10 17:30:30 PDT 2003


RDB wrote:
> I'm curious whether anyone on this list has found real-world circumstances
> in which either semi-polling mode or the net.inet.tcp.inflight_enable
> setting improved performance, and if so what the circumstances were.

Receiver livelock.  However, it's not as good as full LRP.

To see this, put a Gigabit card into a 32bit PCI @ 33MHz, and
drive it full tilt from client machines.

Watch as your data peaks, and then stops coming in at all, because
all your mbufs are used up, and all your PCI bandwidth is being
consumed by DMA's of incoming packets, and you have no cycles left
over to do any work.

Then go read the Jeffrey Mogul DECWRL paper from 1991, and be enlightened.

-- Terry



More information about the freebsd-performance mailing list