IP Switches

[Bdale Garbee <bdale@col.hp.com>]

>    Your question's could be answered many ways depending on how much
> money you want to throw at it. On the high end maybe a Grace card or
> standalone and on the low side a cheap 286 with NRS to TNC's.

Realistically, if you are starting from scratch, (as I am), a Grace standalone
configuration is apt to end up being very cost competitive with an AT-based
switch.  I have been talking to the folks at Grace, and have knowledge of some
future accessory and ehancement plans that will make the Grace solution even
more desireable.  I'd suggest strongly stopping by the Grace booth at Dayton,
or contacting them for new product info shortly afterwards.  Also, note that
the name 'Grace' is apt to change soon due to a legal entanglement.

I spent a little time with a spreadsheet the other day looking at what it
costs to build a multi-port switch.  I happened to be interested in 4 ports,
we'll talk about speed later.  For simplicity, I'll start with 4 1200 baud
ports.  I took the best local prices on a 20Mhz 286 motherboard with 1 meg
of 0-wait RAM, added a floppy-only card, floppy drive, and a 4-channel serial
card with 16550's and added them up.  I left out the case and power supply
since the Grace card is shipped bare (though I believe the simple 5V at less
than 1A needed by the Grace card and a case would be cheaper than an AT case
and supply).  It came to about $495.  Obviously, if you have some parts lying
around, you can reduce this amount, but not repeatedly, and I don't have the
parts lying around.  I chose a 20Mhz 286 because it's faster and cheaper for
NOS than a 16Mhz SX system.

A Grace 5-port standalone card, with RS-232 fanout card, lists for $695.  That
is about $200 more.  Seems like a lot, at first, until you think about the
differences.  The Grace card is *fast*.  The upper limit of the current NOS
on the card hasn't been found because of the difficulty of getting enough
data generators in the right place at the right time, but Don N4PCR reports
running one card with 3 full-duplex 56kb ports and 2 full-duplex 9600 baud
ports, another card with 1 2Mbps full-duplex port and 1 9600 baud port, and
in neither case was the 302 anywhere near running out of gas.  The Motorola
docs say that a 16.67Mhz part (what they're using) should be able to handle
1 port at T1 (1.544Mbps) rate, and 2 ports at 128kbps, all full duplex, with
enough spare processing time to run the two 8530 ports on the Grace card at
9600 and 1200 baud full-duplex, respectively...  The Motorola specs are for
X.25, but Don pointed out that IP switching in their NOS requires similar
overhead to X.25.

Continuing to assume 4 ports at 1200 baud, I worked out a price for a TNC,
ROM, radio, feedline, and antenna at around $400.  I took the Kantronics DVR2-2
price as a good medium between hamfest-cheapo radios and new mobile rigs, both
of which are used in various places.  I assumed 100' of 9913 coax to the
antenna, because that's what we typically use.  For 4 ports, that works out
to $1600.

So, a switch site based on the AT switch is about $2100, and a switch site
based on a Grace card is about $2300.  Using the Grace switch adds about 10%
to the cost of the site, with a spare port left over.  What do we really have?
With the Grace card-based switch, we have a platform with a stable version of
NOS, optimized for the task at hand, with *lots* of room for performance
growth in the future.  With the AT-based switch, we can keep up with the 4
1200 baud ports, and with more money spent, could add faster interface cards
at the expense of overruns on the 16550's.

> Overruns need not be of great
> concern unless they are numerous. An overrun would simply cause a retry
> in the NRS circuit. Something that already happens in the typical multi-node
> netrom connection and probably much more often.

While this may be true, it still isn't at all good.  We may just have
more feeble than average RF paths out here, but retries that happen for
reasons other than thermal noise are so costly on our NET/Wrong "network" as
to make operation tenuous, at best.  After doing the math, I decided that more
money spent on the switch hardware was such a small increase in the overall
site cost that it was worth doing to just not ever have to worry about this
kind of problem again.

> It is not a Grace card running
> 56Kbit ports, but then I am not Fort Knox either. If you are independenly
> rich, don't have a wife or have group support then you could look at the
> high end approach.

I guess the point of this (long, I realize) note is that I'm not sure you are
being fair.

> My goal was to make something work with available
> materials in the best way I could. The 9600 baud (radio) multi-NRS NOS
> switch was the best way for me.

If you had some of the pieces, and are only building one switch, then I won't
argue with you at all.  Problem is, we *know* we need 4 or more switches just
to cover the Colorado front range... and most of our AT/XT pieces are in use.

> I have never tried 19200 NRS, but with
> the radio running at 9600 or less it is not clear that there would be
> any advantage to it.

There is a *big* performance penalty you have to pay for the "extra hop"
induced by the serial port to the TNC when the speed of the asynch line isn't
substantially greater than the speed of the radio port.  It shows up as
increase in the propagation delay on an unloaded channel, and gets much worse
as the data rate increases.  The fact that the Grace card doesn't really need
TNC's offboard, just modems, means you could bring the cost of a Grace switch
installation even closer to the cost of the AT clone installation, and do away
with the overhead.  You could, of course, do this with HDLC cards for the AT
too, but the cost per port is about the same as the TNC's.

The original question was about an 8-port switch.  The Grace standalone
card has 5 ports.  But, you can run the PC plug-in card standalone by
providing power on the edge connector and ignoring the PC interface.  The
plug-in card is $795 list, which would make the 10-port combination of
the plug-in and standalone daughter cards cost about $1500.  This is more
than an AT configuration, but with site costs of about $3800 for an AT-based
switch with 8 ports of 16550, and $4700 for the Grace-based solution.  This
is about 24% more.  This is indeed a lot of money.  But, to be fair, since
you only really need modems and not TNC's on each port, and if you're doing
more than one 8-port switch you're doing enough ports to justify scrounging
or building modems, the real cost of a Grace site might be more on the order
of $4000, which is only 5% more than an AT solution!  In fact, it gets even
cheaper if you can configure your network to not have octopus-nodes, so you
can stick with a single 5-port switch per site... that's what we're doing
here.  We'll have a mixture of 1-2Mbps 10Ghz links, 56kb links, and 9600 and
1200 access channels.  The Grace cards are ideal interfaces for all of these
"modems".

I've bought a Grace standalone card, and convinced two friends to each buy one
too.  I will soon be focussed on adding functionality to their port of NOS,
particularly looking at the routing and network management issues.  The PS-186
is, I now believe, really dead, and the DE is too port-limited for many uses.
AEA was talking about a list price for the PS-186 very close to the cost of
the Grace card, anyway, and I don't believe it will ever run NOS as fast, for
a variety of reasons, most notably the way the 68302 DMA dovetails with the
mbuf design in NOS...

I have had NOSINABOX limping along on the DE since Dayton, but it is port-
limited, and I have decided that our local needs would be better served by my
focussing on taking the already quite pleasant Grace NOSINABOX and adding more
functionality to it.  I've put some info in my yet-to-be-published PSR column
this time around about all this, but I haven't said much publicly until now.

I put a request in my PSR column asking folks to help me with a
survey of what it really costs to build a packet facility, be it a PBBS, a
digipeater, or a multi-port node.  I'd like to extend the same invitation to
the readership of tcp-group.  What does it really cost, anyway?  And how are
the costs broken down?  I've taken a stab based on my experience here in
southeast Colorado... are my numbers anything like yours?  Seems like this is
something worth talking about, so we can all get the most value from our
always-limited money.  You see, I'm married, I'm not independently wealthy,
and packet isn't my only cash-sink, either...  /o\  I just have this hang-up
about value, and the Grace card is to me clearly the best value in packet
radio today.

Bdale