Robert E. Seastrom writes:
From: Alexis Rosen <alexis@panix.com>
[...]
This isn't clear to me. Why do you assume a ZIP is likely to be more
reliable that a hard disk? ZIPs haven't been around long enough to be
sure of this, and HDs are pretty reliable these days.I think his plan was to boot ramdisk unix from it and then to spin the
unit down. Spun-down units are fairly reliable, and besides I think
the main thrust here was to replace the drive with something that
could be swapped easily for upgrades.
That's more easily done using Vixie's scheme. *My* main thrust is a system
with no moving parts (assuming that that's a good way to achieve higher
reliability). Of course, both goals are desireable.
Of course I'm not saying that I *Want* to use an HD in this situation; flash
is clearly a big win. But I don't see how using a floppy or ZIP improves
wins.With cold-convenient-swappable IDE drawers that let even a
kindergartener swap out an IDE hard drive and high quality 100mb hard
drives available for like $50 (at this point you're probably paying
more for the snazzy mounting kit than you are for the drive), I
daresay the Zip and flash solutions are far too expensive for what
they buy you. Take a look at the MTBF on your hard drives and then
look at the MTBF on power supplies and floppies, and gee... Alexis is
dead on here. Pay more, get less...
I think that realistically, finding a stable supply of anything smaller than
1GB is unlikely, unless you're buying large quantities. More to the point,
disks for less than $100 are unlikely. But it's not the money that's key
here. One service call to a lights-out facility costs lots more than that.
Another solution occurs to me. Use dual-anything (floppies, HDs, flash gizmos).
Pay someone to modify the BIOS in this small but important way: have it
alternate which device it boots from. Let the machine write check a state
file kept somewhere else each time it boots. If the same device has booted
twice in a row, there's a problem and you notify the NOC that a device has
failed. Otherwise rewrite the state file.
There's another better solution. It would take a little bit more work but
it would be infinitely more useful: Build an ISA card that looks like an
MDA adapter, but which sends output to a VT100. (A simple algorithm will
produce good results here even when the cursor jumps all over the place.)
You also need a gizmo (one exists alread) that makes serial input look like
PC keyboard scan codes. And lastly, you need to special-case a long-break
so that it causes the ISA card to reset the machine.
All of a sudden you've got a working remote console device. This is an
extraordinarily useful gizmo, and I'll bet you could sell a million of
them. (Everyone who uses Linux, FreeBSD, or NetBSD as a server would buy
it, as would the millions of poor souls who run Novell, SCO, etc.) I'd
buy fifty of them at $100 apiece, and at that price you'd probably see
a gross margin of 300%, even in small quantities, and including the keyboard
gizmo. (Now all we need is a better word than "gizmpo" and we've got a
real product. 
Anyone interested in some hardware design work?
/a
