Going to be One Of Those Days

[Roy S. Rapoport]

I hesitate to enter into this discussion because I suspect we'll be coming
up on people's religious biases (including my own -- I'm a Solaris guy, and
I know exactly the pleasure that is working on Sparc platforms).
Nonetheless, I'll make a few comments.

On Tue, Sep 14, 2004 at 10:59:42PM -0700, Guy B. Purcell wrote:
> I'm not so sure.  I'd _love_ to see a real study done that takes into 
> account all the relevant factors (or as close to that as is 
> practical)--including extra HW required for remote reboot, and extra 
> sysadmin time to assemble the ultra-cheap build-it-yourself boxes, and 
> to replace cheap HW as it dies under stress, etc..

Two notes:

PCs can serve in, generally, two capacities:  Desktops and servers.  I've
seen companies move from Sparcs on the desk to PCs on the desk and more
centralized UNIX systems at the center.  In that case, things like remote
booting and serial communications become less critical.  Putting PCs at the
client side also allows for some cool things -- current workplace issues
notebooks to EVERYONE, which allows anyone who wants to take their work
home.  From work's perspective, this is a good thing.  We won't get into
Tadpole's Sparc notebook offerings :)

On the server side, we can again differentiate between critical one-off
servers and undifferentiated server farms.  In the case of undifferentiated
server farms, issues like serial communication and remote unwedging become
a lot less critical.  You need X machines to serve something; you've got X+N.
If Y (<N) machines are having problems, deal with it later, and at your
convenience.  

Sure, systems will go wrong, and parts will be replaced.  If you're doing
parts maintenance often enough, then the company is insane if it has its
senior (or even mid-level) engineers doing parts replacement -- that's what
hardware monkeys are for (I'm ignoring situations like startups where
people's time is considered more inexpensive and you often have people in
senior position who _like_ tinkering with hardware); 

> At $CURRENT_JOB, we "upgraded" from sturdy-but-old SPARC boxes to Intel 
> HW from Sun.  The HW wasn't any cheaper than similar SPARC boxes 
> (V60x's vs. V210's), and it has held up well under load so far, but 
> also has required almost $1,000 extra in manageable power strips, the 
> OS support ($BOSS requires OS support) fees are more than they were for 
> Solaris, and I had to add the optional bits (second CPU, second disk, 
> more RAM--stuff my VAR should have done, if I had a decent one; don't 
> ask--we can't switch) to each one (a significant time cost for around 
> 40 boxes).
> 
> We originally went with much cheaper Intel HW, but it broke seriously 
> under load, so decided that "you get what you pay for" is somewhat 
> close to correct.  I have this uncomfortable feeling that that adage is 
> deeper than typically interpreted, and that "cheap" HW ends up costing 
> the same as or more than the "expensive" stuff in the long run.

I spent about an hour last night trying to figure out why a drive I had put
into a PC running RHEL3 caused my bootup to get horked.  Answer:  Because
it came from another system and had some of the same volume labels as my
main drive.  Being used to "put it in, make sure it's not an existing SCSI
ID; reboot with -r; format; newfs," this was a ... lackluster experience.
In fact, I'll go further and say it was lame.

So hey, _I'm_ not going to come out and say that PC hardware is the way to
go.  

I will note that buying PCs from Sun seems ... well, it seems a little
silly to me, if your goal is to save money.  Hell, buying _anything_ from
Sun seems a little silly to me, if your goal is to save money :).  

My perception is, and I'd want to see hard numbers if you intend to
disprove it, that for most workloads, you can get cheaper hardware, from an
initial investment perspective (we're not talking TCO here), to serve the
needs on PC rather than the premium UNIX platforms.  The parts costs for
replacing broken hardware (we probably want to address service offerings,
but it seems that while Sun continues to request you bend over every time
you negotiate a service contract with them, their quality has dropped, from
what I hear) seem consistently lower.  So the issue then is time.

Inexpensive, decent Linux engineers seem to be easier to find than
inexpensive, decent Sun engineers.  

I do know people who use PCs in a corporate environment and who seem to
have done the financial legwork to prove to themselves that it was the
right way to do things.  My favorite search engine, for one.  

-roy