response 56 of 181:

Oct 15 18:41 UTC 2001

We do say, "..things change constantly..". At issue is the meaning
of "constant". The sense in "change is the only constant" is that
"change is constantly occurring". Dictionary definitions include
"long continuing or continually recurring", which change certainly
can. 

response 58 of 181:

Oct 15 21:40 UTC 2001

Re #49: Yes, e^x *is* its own derivitive. That's what's special about the
constant 'e'.

response 60 of 181:

Oct 16 00:42 UTC 2001

resp:56 - "long continuing or continually recurring" can indeed 
describe real-world change.  But it also describes real-world non-
change in many circumstances.  So change remains not the only constant, 
regardless of the definition of "constant".

response 62 of 181:

Oct 16 23:31 UTC 2001

Re #39: What data do we have that says that grex's user base is
declining, and is grex in any danger of extinction?

Re #42: I suspect that buying a server class rackmount case, power
supply, motherboard, sca drive bays, possibly floppy and or cd-rom
drive as a single package, and then adding ram, disks, and processors
may actually be what makes sense, rather than getting individual
pieces.

server class machines these days usually have scsi, video, and
ethernet on the motherboard, such that you usually end up not needing
pci cards at all, especially in the 1U and 2U servers.

A gig of RAM is like $400 or something, last I checked.  More than
that gets expensive quickly because you generally have to buy 256MB or
larger DIMMs.

dell and compaq have their own unique rackmounting hardware, which
makes it difficult to mix their hardware with other vendors' hardware,
and you generally want things like ethernet switches and upses in the
rack, too...

Re #43: if it compiles faster, the vandals will finish seeing how
futile their efforts are faster, and with less effect of slowing down
the system.

Re #47: I have a 60 mhz pentium with a 120MB hard drive acting as a
kerberos server.  It works well enough that I'm not likely to worry
about finding better hardware for it anytime soon.  So I think that
claiming that two year old hardware is obsolete is misleading.  And my
mail server literally is about two years old (although its hard disk
and power supply have needed to be replaced, and I added RAM a few
months ago), and if I replace it, it's going to be more because I want
SCA drive bays and an emergency management port than that the current
hardware is inadaquate.

Re #50: I'm not really convinced that the NetBSD folks get bored with
maintaining software for older platforms.  And in general, once they
get it to work they generally don't do a whole lot to break it.

I think I've experienced far more problems from stuff that was too new
(raidframe in netbsd 1.4.x, a usb cf reader on 1.5.2) than I've had on
stuff that was too old (admittedly, I have a solbourne that netbsd
simply doesn't support, but it never has, either).

Also, just because the modem lines are seeing declining amount of use
does not mean that we should ignore the possibility of improving the
quality of the existing lines.

Re #52: RAID mirroring means that whomever has to deal with the failed
disk has a lot less hassle to deal with when they fix the problem, and
a disk failure isn't nearly as likely to completely take the system
down, requiring immediate attention.  Assuming a good raid
implementation (not that I'm convinced that any exist), the total work
is likely to be reduced by using raid mirroring.

response 64 of 181:

Oct 17 04:28 UTC 2001

(And enough staff to do the maintaining, and someone to pay the bills and
collect the money.)

response 68 of 181:

Oct 18 00:55 UTC 2001

I'm certainly leaning toward an Intel machine, but I'm not a concensus.  I'd
be perfectly OK with the Sun option, and I know some staff members like Sun
a lot.  Holding off shipping things certainly wouldn't hurt.

response 70 of 181:

Oct 18 01:37 UTC 2001

Yeah, I can look after a server over here.

response 72 of 181:

Oct 18 03:40 UTC 2001

I'm still a little vague about what services we would distribute.

response 74 of 181:

Oct 18 04:16 UTC 2001

Jan, doesn't putting mail operations on a separate machine count as 
distributing services?

response 76 of 181:

Oct 18 14:23 UTC 2001

   Finally, I've caught up with this idea.

   I know that the end for dear old SunOS and Grex is near.  I always
thought that the disk size (we can't use more than a 16G disk) and
things like disk quotas were going to be the finishing points for
Grex on SunOS, but hearing that mysql and other software won't easily
(or at all?) port is bad news.  Yes, we have to upgrade.  It sounds
like we're going to have to do it before too long.

   Right now I see one software option and two hardware choices for
us.  The only software I want to run is OpenBSD.  Why?  Because Grex
gets hit by every vandal around (I'm not sure that isn't literally 
true), and we need something that is *rock solid*.  Nothing less
will do, and we're kidding outselves if we don't put that as prio
#1.  Number one.  As for hardware, the choices are i386 and sparc64.
I maintain a server which is a Dell 1300 PowerEdge running OpenBSD
and its solid.  So solid in fact, that in two years of uptime it's
only been rebooted because of humans depriving it of power.  It's
that good.

   The staff time factor to deal with it all is a problem, but if
we look back at all the efforts we've done over the last five years,
especially emergencies, disks become our biggest problem.  We've had
three real disk disasters which we've managed to get around (once
with the SMD disks back when we were at the warehouse, and twice
while in the Pumpkin), so I'd have to argue that the kind of hardware
we use is secondary to the fact that Grex kills disks, and that has
been--and is likely to continued to be--our biggest problem.

   i386 stuff works.  My open experience with polisci has shown me
that with *GOOD* hardware, which I consider Dell to be, the system
can be very very reliable.  Still, I would prefer Sun equipment, for
a couple of reasons.  The first is the quality of the larger systems
intended for "server" use is still fantastic, and I'd pit them against
i386 stuff in any hardware scrutiny contest.  But the real reason, the
one which matters for Grex is that such hardware was designed for the
running of many things at once.  This is an important difference, which
folks who've only used i386 hardware usually don't see.  Looking at
our current hardware, we have four 40MHz CPUs, yet even with 70 users
on, our speed while slow is still OK.  I have seen 350 Mhz i386 systems
which have been slower than Grex, which have had (at least in theory)
a faster CPU and faster disks.  Still, it ran slower, and the reason
for that is simple: Pentii aren't designed from the ground up to be a
multi-process machine; not really.  Sure, they can do this, but not
as well as the Sun-4/670 does, because it has special hardware
integrated with the CPU called hardware contexts, which help run
many processes at once.  The difference might be akin to a racing
hound and a collie: both can run, but one is more optimized for
speed.

   I think that we'd be better off with Sun sparc64 equipment, and
once we made the jump to it, we'd be in another whole class of
hardware platform, such that we'd be able to move to faster stuff
without software changes.  In the sparc64 lineup are more server
class machines, which is exacty what Grex needs.  Yes, a Pentii
type machine is "faster" in terms of the CPU, but Grex needs
throughput, which isn't the same.  We'd be better off with Sun
on that score.

   I'm going to have to finish this later on today, as krj is
about to get me so we can get to work.  I'll finish this today
however.

response 78 of 181:

Oct 18 16:46 UTC 2001

Sun machines are generally going to be more expensive, but they're 
generally of sturdier construction.

However, I think we've been offered some used Sun equipment; that might 
be worth considering.

response 80 of 181:

Oct 18 23:17 UTC 2001

STeve, I'm not really sure that Sun has the hardware advantage you think it
does (in terms of multi-user performance).  For example, look at the top
performers of the TPC-C benchmark (I know that it's silly to compare Grex to
$10M machines simulating a warehouse order entry operation, but it does
illustrate hardware performance across different architectures).  You'll see
that amongst clustered solutions, the i386 machines do quite well.

As for price/performance, the top-10 systems are all i386 based. Given
the level of performance required, an i386 system could well be sufficient.
(Consider that a 1.6 GHz Pentium 4 system -- configured with 512 MB and
with a much faster bus than older i386 architectures -- can be purchased for
about $900 (going rackmount and adding SCSI drives will increase the cost.)