It's possible for two processes attempting to start diald on the same
interface at the same time to interfere with one another.  One cause
of this is that the filenames in /tmp are unique only by interface.  I
haven't thought through the causes, effects, or solutions.  This
problem never exhibits itself when diald is started with the standard
system initialization methods.

There's no way to tell which dctrl goes with which interface.  I tried
-title as a dctrl argument, but no go.

The output of the connection script displayed by dctrl when
VERBOSE_LOGGING is off sometimes lags behind what is actually being
done by the modem.  This is because the modem often does not generate
the line feeds that flush the data in the pipes.  The workaround is to
turn VERBOSE_LOGGING on to debug connection problems.

You will have to write a /etc/ppp/ip-up script to control route
management when using more than one diald interface.  There's no
provision in the RedHat ppp setup for establishment of routes when the
ppp interface comes up.  This makes it more difficult when managing
multiple interfaces.  You will also need to use the ADDROUTE
configuration directive to add routes to the additional (non-default
route) proxy slip interfaces established by diald.  This can be done
by pointing ADDROUTE to the redhat static route adder, as in the
example in the README, and using the Redhat control panel to add
routes to the interfaces.  Because there is no provision for a metric
in the redhat static routes, the /etc/ppp/ip-up script will have to
remove the static route established by diald to the proxy serial
interface, and a ip-down script will have to reestablish it.

The use of Martian IP addresses for the diald's proxy slip interfaces
causes problems when it is a tcp socket that trigger's diald's dialing
of the phone.  Such a tcp socket will fail to open because, among
other routing problems, the Martian address at the local end of the
socket will never have packets routed to it by the internet.  This
problem does not exist when the local ip addresse is not dynamically
assigned by the ISP.  Typically, this is not a problem because the
first packets generated by applications are usually DNS udp packets,
but when a local nameserver is used the name may be resolved locally
and the first packets destined for the internet may be tcp packets.
Thus, the application which raised the link may fail.  I suspect that
clever application of masquerading and dynamic route establishment may
work around this problem.

There seem to be lots of fifos left lying about /tmp by dctrl, or
something.

Read the mgetty documentation for an explanation of why you need to
use a ttyS<n> device and why all programs accessing a particular
serial line device must use the same name for the device.

Some programs, notably sendmail and named, require net access when
they start up.  This can cause the phone to dial.

Support for clones are minimal.  There should be:
/etc/sysconfig/network-config/dialdcfg-<iface>-<clone>
/etc/diald/diald-<iface>-<clone>.conf
/etc/diald/diald-<iface>-<clone>.conf.m4
but there arn't.  Also need clone support in start-dctrl.  I don't
like clones anyhow because I think they will have to be "undone" to
get the proper modeling between configuration entities and real-world
entities.
