HP-UX TCP/IP Performance White Paper, March 2008

73
tcp_keepalive_interval:
Interval for sending keep-alive probes.
If any activity has occurred on the connection or if there is any
unacknowledged data when the time-out period expires, the timer is
simply restarted. If the remote system has crashed and rebooted,
it will presumably know nothing about this connection, and it will
issue an RST in response to the ACK. Receipt of the RST will
terminate the connection.
If the keepalive packet is not ACK'd by the remote TCP, the normal
retransmission time-out will eventually exceed threshold R2, and
the connection will be terminated.
With this keepalive behavior, a connection can time-out and
terminate without actually receiving an RST from the remote TCP.
[10000, 10*24*3600000] Default: 2 * 3600000 (2 hours)
These keepalives for an established connection will only be enabled if the application uses a
setsockopt()/t_optmgt() call to enable keepalives (SO_KEEPALIVE). It is not enabled otherwise.
This differs from the behavior of tcp_keepalive_detached_interval. A large number of idle
connections, with keepalives enabled, could generate more keepalive traffic than real traffic. They could
also keep on-demand/dial-up links open unnecessarily.
tcp_largest_anon_port:
Largest anonymous port number to use. [1024, 65535]
Default: 65535
It is unlikely that this will ever need to be changed. The only reason would be to provide some similarity to
anonymous (aka ephemeral) port number assignments of older BSDish stacks. There are no checks made to
ensure that tcp_largest_anon_port is set to a sane value wrt to the smallest anonymous port number.
Setting tcp_largest_anon_port to a value smaller than tcp_smallest_anon_port will lead to
undefined behavior.
tcp_recv_hiwater_def:
The maximum size for the receive window. [4096,-]
Default: 32768 bytes
This can also be thought-of as the default receive socket buffer (aka SO_RCVBUF) size. It is used for any
connection using an interface that is NOT marked as either a "long, fat pipe" (LFP) or a "long, narrow
pipe" (LNP).
One of the "limits" to the performance of a TCP connection is based on the relationship between the TCP
window (W) and the round-trip-time (RTT) of the connection. Basically, a sending TCP cannot send more
than one window's worth of data before it must stop and wait for a "window update" from the remote. The
soonest a window update can arrive is one RTT. From this we have the function:
Throughput <= W/RTT