Re: ssh, HPN extension and TCP auto-tuning

[Brian Buhrow <buhrow%nfbcal.org@localhost>]

	Hello.   I'm not sure if the kern.sbmax value needs to be as large as
the window size since, presumably, on the receive side, the data is sent to
the application long before the socket buffer gets full.  And, on the
transmit side, I think the data is passed into the stack as MTU-sized mbufs
and stored as a chain, rather than as one large socket buffer.  But, I
could be completely off onn this.  However, if I look at slow TCP
connections which fill up, the send-queue size seems to reflect the window
size, rather than the socket buffer size.  I'll pay more attention to this
in the future and see if I've missed that obvius adjustment.

	I can tell you, however, that with those tuning parameters in place, I
can fill a 100 mbits/sec pipe with ssh traffic.  Before those changes, I
could not.
-thanks
-Brian

On Sep 21, 11:25am, Havard Eidnes wrote:
} Subject: Re: ssh, HPN extension and TCP auto-tuning
} Hi,
} 
} > </etc/sysctl.conf>
} > # Improves TCP performance significantly with ssh.
} > net.inet.tcp.recvbuf_auto=3D1
} > net.inet.tcp.sendbuf_auto=3D1
} > net.inet.tcp.sendbuf_max=3D16777216
} > net.inet.tcp.recvbuf_max=3D16777216
} 
} Thanks for the suggestions, and I've done some initial
} adjustments with beneficial results.  I was a bit more
} conservative and went for a 1MB sendbuf_max / recvbuf_max.
} 
} One thing I didn't see was any corresponding adjustment of
} kern.sbmax; doesn't it also need to be as large as you want the
} TCP window to be able to grow?
} 
} Best regards,
} 
} - H=E5vard
>-- End of excerpt from Havard Eidnes