net.inet.tcp.inflight_enable

[Noritoshi Demizu]

> Sorry, actually, I always set net.inet.tcp.inflight_enable to zero
> both on DragonFlyBSD and FreeBSD in my experiences (I know it is zero
> on DragonFlyBSD by default, but I want to make it sure), because
> unfortunately it reduces throughput in my experiences in an unexpected
> way.

I'd like to show an example where TCP throughput was reduced by
setting net.inet.tcp.inflight_enable=1.

  http://www.demizu.org/~noritosi/memo/2005/0706/

The first graph shows a TCP behavior when net.inet.tcp.inflight_enable=0.
Since congestion window grows exponentially, it took 5 RTTs (about 100ms)
to send 64KB data.

The second graph shows a TCP behavior when net.inet.tcp.inflight_enable=1.
Since congestion window is almost unchanged, it took 14 RTTs (about 300ms)
to send 64KB data.

When I sent larger data, the difference became much larger.
(e.g., 167ms and 1330ms when I sent 512KB data.  RTT=20ms)

(Note: according to the graphs,
 "cwnd = 3 * 2^RTT" vs. "cwnd = 3 + RTT/N (N is about 10?)")

Regards,
Noritoshi Demizu