latency: make a TCP connection to a remote host and measure "keystroke"
	latency. Print out statistics and a histogram of data.

Usage:
	latency [<options>] <host>

Options:

	-r <prompt>	Use "rlogin" instead of "echo" service.
			<prompt> matches the remote host cmd line prompt.

			Note: set <prompt> to "." to match trimmed hostname.
			(e.g. "foo.bar.com" gets trimmed to "foo")

	-s <service>	Use <service> for a dialog with remote host.
			Currently supported:
				echo|rlogin|smtp|ftp|nntp|annex|\d+
			(Hack: annex only works for vanilla Xylogics.
			       \d+ means a port number redirected to
			       the echo (7) port on some machine.)

	-b <size>	For echo(7) service send text of <size> bytes.
			and a newline (default: 1). Note that -l
			flag will send an extra 5 bytes indicating the
			packet number.

	-C <char>	Use <char> as the character to send instead of "X"

	-c <count>	Measure <count> keystrokes. (default: 1000)

	-f <count>	Print out info every <count> keystrokes (otherwise
			the printing frequency is determined dynamically).

	-k <count>	Skip the first <count> packets in the averages.

	-l		Loop forever.

	-R <n>		Refresh running min/max/ave stats every <n> seconds.

	-S		Scroll lines each refresh interval.

	-X <code>	Extended scrolling.  When <code> yields true, print
			a newline + timestamp.	You can learn the available
			variable names, etc, from reading this script. Look
			for "ExtendedScrolling". IF ... THEN ... END ...
			are described there as well with examples.

			Also, -X <n> is shorthand for times <n> times greater
			than the running average of times.

	-i <time>	Sleep this (floating point) amount of time between
			keystrokes. (default: 0.0)

	-q		Be quiet while collecting the data.

	-nc		Do not read ~/.latencyrc or ./.latencyrc files.

	-n <n>		Make histogram have <n> bins. (default: 78)
			Setting this will make the X-Y plot output.
			(Hack: unless <n> starts with "a", e.g. "a40", will
			 make ascii plot only and with 40 bins)

	-T t1,t2	Plot only times t with t1 < t < t2, e.g. -T 0.01,0.50

	-H <f>		On the ascii plot, only show data between zero and
			fraction <f> of maximum height. e.g. -H 0.01

	-t <mode>	Method to gettimeofday(). <mode> can be
			'external' for an external C program or
			'syscall' to use the perl syscall function.
			Prefix mode with 'test-' to display some clock
			timing tests.
			Default: syscall

	-p		Output X-Y plot histogram to stdout.

	-np		No plots at all.

	-h		Print out this help.

	-K		Do not remove the timer program at end of run.

			The timer program is attempted to be kept around
			in /tmp/timeprog by default. -K makes sure it is
			not removed whatever its name is.

Notes:
	Remote host must not deny "echo" service to you.

	Instead of -s you can use "host:service", e.g. foobar:4321

	With -r, remote host must use hosts.equiv or .rhosts to let
	you rlogin w/o a password.

	With -p, stderr will still get the lovely ascii plot.

	The prompt match is actually "\s<prompt>\S*\s", the \S* allows
	chars appended to the prompt, e.g. "hostname% "
	
	When doing rlogin, using -r 'hostname\[' is often useful to
	avoid matching the hostname too early.

	When doing rlogin or annex if the hostname contains a space,
	then the hostname is really a command to attach to instead.
	If hostname is like /dev/ttypa or /dev/ttypa:23452 then the
	special file is connected to.  For the second case the numerical
	pid of the process follows the ":" so it can be killed when done.

	~/.latencyrc file contains perl code to be sourced at beginning.
	See the "Defaults" section at the top of the script for ideas
	about what can be set.

	Joe says the only way to go is: latency -l -X 4 -np <machine>

Comparison with "ping":
	+ Pings faster (like ping -f, but always gets an ACK).
	+ Measures the interactive latency caused by dropped packets.
	+ Nifty histograms/plots of ping times and ave & standard deviation.
	+ Probes the userland component of the connection via TCP.
	+ Measures "real" user response via rlogin.
	+ Scolling can be triggered by events of interest.
	- Hi-res time measurement not very robust.
	- Requires "echo", "rlogin" or other services provided by remote host.
	- Only sends 1 or 2 bytes of actual data per packet.
	- Depends on perl and possibly cc(!)