Re: [tlug] Who do you recommend for Business Desktop?

[Jonathan Byrne <jq@example.com>]

This may be a bigger answer than you were really looking for :-)

On Sun, Feb 29, 2004 at 05:14:25PM +0900, Thomas Savarimuthu wrote:

>I wonder who can be the best in the following list in terms of performance, 
>cost and support. Dell, Compaq/HP,IBM, Fujitsu, NEC, Gateway, Sotec(?)...

WRT performance you will:

1) Find no significant variation between any of those, between a white box
machine from a local shop, or from a white box business workstation you
build yourself;

2) Have more than you could possibly need.  A PIII-1000 is more than 
adequate to power a business workstation, yet the Dell they issued
me at work has a PIV-2400 in it.  Absolute overkill.  I know of a
Duron 900 box that runs XP at a perfectly acceptable speed.  Heck,
my workstation at GOL had a PIII-450 or 500 in it, and that was quite
acceptable running Gnome on RH 7.3.  Brett might still be using it
now, he could tell you how it's doing :-)

I haven't seen or opened a Fujitsu in a long time, but I once
worked for a company that had a bunch of Fujitsu FM/V machines in the
late 1990s.  We had all sorts of issues with them.  Almost every one
had a quirk of some kind.

Dell machines are well made and have a very clean layout inside.  I've
taken several of them apart and it kind of reminded me of working on
a Sparcstation.  I expect you'd find similar inside of any of the other
machines.  Among the major vendors, just get the best deal you can.

Also, if it's an option, take a look at some local whitebox suppliers
and see if they can offer you a better price than the majors.  If they
can, you might want to consider going with one of them.  Chances are good
that you will get a machine with more standard parts (read "easier to
upgrade") than if you buy a low-cost business machine from one of the
big three.  You might also get better support, although nothing is ever
guaranteed anywhere.  However, a person who buys 20 or 30 workstations
from a local business is a much bigger fish than a person who busy 20 or
30 workstations from one of the big suppliers.

As someone else mentioned, building your own is an option if you have the
time.  You aren't likely to save money this way and might even spend more,
but as noted, you'll have complete creative control over what goes into
every machine and they will have maximum upgradeability down the line.

Either a local whitebox machine or a build-your-own will offer much
better upgradeablity than a major brand business machine because of
standard components and cases versus possible custom components.  Cheap
business machines are basically not designed to be upgraded.  They
are designed to be replaced every few years when the company does a forklift
upgrade of all its machines.

Finally, if these will be Linux boxes, one last option is to build
diskless workstations and netboot them.  All you need then is a motherboard,
cheap video card, and network card that you can put a boot ROM into
for the netboot.  You might be able to get an all-in-one motherboard that
covers all of those bases.


A couple years ago, somebody (I think it was Mauro) did a presentation on
the Linux Terminal Server project at a TLUG meeting.  If I were setting up
an office with Linux machines, I would take a good look at this.  It saves
you money on hardware costs, improves security, simplifies backups,
and a user's desktop environment will follow her to any workstation she uses.
All she has to do is log in and there it is.

This is practically heresy, but even a sysadmin can use a diskless
workstation successfully.  The only counter-arguments for a full 
computer for sysadmins and developers is that they can maintain it
themselves (well, sysadmins can; you'd be surprised how little some
developers know about computers) and it offloads the disk work of compiling
onto their workstation.  If I had a lot of developers and gave them 
diskless workstations, I'd also give them a high-powered multi-CPU dev
box which was separate from the main server.

That leads to the one thing in a diskless workstation shop that
does tend to be of higher cost.  The central server.  It should 
have redundant hot-swappable components to the greatest extent you can
afford.  This should include the power supplies, which are the things
most likely to fail in most computers.  All disks should use hardware
RAID.  Not baby ATA RAID like Promise boards, which are really hardware-
assisted RAID in which all of the heavy lifting is done by the driver
(which makes them software RAID, basically; the hardware just does some
management functions ).  For ATA, 3Ware makes real RAID controllers and
they are well-supported under Linux.  For SCSI, it's hard to go wrong
with Adaptec, although they are not the only vendor.  They do, however,
have a very long Linux history.  These 3Ware and Adaptec boards are expensive
(hundreds of dollars), but on a mission-critical server, full hardware-
accelerated RAID is what you want.

For areas of the main server which cannot be made redundant (such as the
motherboard), keep a spare on-site.  Consider using hot-swappable disks
(this means SCSI).  If that's too expensive, look for a case with screwless
mounts and keep at least a couple of your spare disks ready to go, with
the mounting rails attached if required, so that all you have to do in
the event of a disk failure is schedule a brief downtime, power down the
machine, pull the bad disk, put in the new disk, and power up again.
Downtime should be less than five minutes.

Label the disks in the machine so you know which one is which, and make
sure you can see those labels clearly when the drives are in the bays.

A multi-CPU machine is a must.  Not only for performance, but because
if one CPU fails in the middle of the business day and you don't have
a spare, you can just remove it and power up again and run at half-speed
until you can get a replacement.  Do, however, keep at least one spare
CPU in reserve, since CPUs do go out of production.  I have
a dual PIII-1000 motherboard with only one CPU in it, and I doubt I'll
ever find a match for that one. 

Put lots of memory in such a machine.  Putting in as much as the 
motherboard will take would not be a bad idea.  Keeping spare memory
on hand is also not a bad idea.  Use premium memory such as Micron
or Crucial.  No bargain-basement stuff.

Big UPSes.  If you have succeeded in getting a machine with redundant
power supplies, plug each of them into an APC 700 or bigger (a 1500
for each is better if it's in your budget), and if at all possible,
plug each of those UPSes into a different circuit.  Cut the breaker
that each circuit is supposed to go to, to make sure the building is
really wired the way they say it is.  A pair of APC 1500s will last
that machine through a pretty serious power outage.  That's expensive,
but if you don't care if your staff can work during an outage, you
can save money by not putting UPSes on any of their diskless workstations
(and most companies don't put UPSes on any staff computers anyway, so
you're probably already doing this).

Also, pay attention to the network infrastructure.  You probably won't
need Gig-E unless this is a pretty good-sized company, and even then you
can probably get adequate performance by just using Gig-E trunks and
100 megabit distribution switches for each floor.  That doesn't 
necessarily mean every switch needs to be a Cisco (although that's
not bad), but on the other hand, if the operation is big enough to
need Gig-E anywhere in the network, then maybe they should be Cisco or
some other brand of enterprise switch.  Whatever switches you use, have
at least one spare around.  If they are cheap switches, keep several
spares.  If one dies, the people hanging on that switch don't work, so
you want to be able to replace it quickly.

Finally, use good network cables.  Cat5e or Cat6 everywhere.  Use
stranded-wire cables (not solid) to connect the workstations to the
switches.  Solid wire does not stand up well to things like bending
around corners, table legs, etc.  Even though only four of the eight
wires in the cable are actually used by Ethernet, Murphy's Law requires
that only those four can break :-)  Of course, keep spare cables around.

While diskless workstations probably won't save quite as much as you
might think on first glance, due to the higher costs of your server
and network infrastructure to support them, they should save some money
upfront when outfitting a new shop or doing a forklift upgrade, and
will save money in running costs, because those workstations will continue
to be cheaper to have and upgrade, and also because they won't need to
be upgraded or repaired as often, because the have so little in the
way of parts.

Finally, for sysadmins, software upgrades are easier because the server
is the only machine that needs an OS upgrade.  For the workstations,
the upgrade just consists of popping in a new ROM firmware on the NIC
if/when needed.  Of course, keep a few spare NICs with ready-to-go ROMs
in them sitting around, too.  On the odd chance that one dies, you can
replace it and have that workstation back up in a few minutes.

Jonathan
-- 
gpg --keyserver pgp.mit.edu --recv-keys ACC46EF9
Key fingerprint = E52E 8153 8F37 74AF C04D  0714 364F 540E ACC4 6EF9
"99 pounds of natural-born goodness, 99 pounds of soul!"

Attachment: signature.asc
Description: Digital signature