[Beowulf] Definition of HPC

[Lux, Jim (337C)]

On 4/16/13 12:50 PM, "Max R. Dechantsreiter" <max at performancejones.com>
wrote:

>...Perhaps HPC should be defined problem-down, rather than hardware-up.
>
>Regarding your "Issue #1:"
>
>There are some _nearly_ EP problems that require very minimal, but
>crucial,
>communication (perhaps at long intervals).
>
>As to your "Issue #2:"
>
>"Owned compute" has some advantages over "rented compute."  In general,
>the
>control one has over one's owned resources enables applications to run
>with
>greater performance.  Some optimizations just demand root access!
>

There's also the "budget bucket" from which the money comes..
For instance, the US Govt doesn't depreciate assets, nor do they borrow
money to buy things (at least not like a business does).  Once you've
acquired something, it's entirely paid for.  If it sits unused with the
power off, that's not "costing" any money from your budget.  In the long
run, Congress wants to see that investments are made wisely and that
you're getting good use out of whatever you spent that money on.  But it's
very different than the "cost recovery" model used in a lot of businesses,
where the users are charged a fee based on a combination of depreciation
and actual operating expense.

It's one reason why Government labs tend to have a lot of 10-15-20 year
old test equipment. With leased equipment, you might as well have the
newer gear, because the lease price effectively pays to replace it every 3
or 4 years.  At a Govt lab, you buy gear on a big project (where it is a
small fraction of the overall cost), and then other projects just inherit
it until it falls apart.  There's a lot of test equipment that was
purchased for Galileo and Cassini that is still being used at JPL, because
it is essentially "free" for the cost of repair and calibration.

Since the industry has gone to shorter life cycles (what with embedded PC
controllers inside the instrument, there's a significant problem with
software support), I've noticed that the "repairability" of equipment is
less.  Partly, too, it's due to higher levels of integration.  IN a lease
environment, this doesn't matter, but it is a big sea change for people
used to the purchase and use for 20 years model.


In any case.. Once I've bought that cluster, and I've got it stashed in
the office next to mine, I can use it or not, essentially for free.
Nobody hassles you about trying to "make best value of the resource". It's
there when you need it.

OTOH, if you want to "buy time" in the cloud, you need to issue a
procurement requisition, which has to go through all the contractual
hoops, and it hits your budget, right now.

So, if you have a sort of speculative idea, and you'd just like to try it
out; with the owned cluster, it's whether you can arm wrestle the other
users to get some time, since presumably YOUR time is already paid for.
With the cloud services, you need to have an budget account to spend from.


Hewing to Hamming's quote "the purpose of computing is insight, not
numbers", personal computing (whether super, HP, HA, or whatever) supports
that insight because it's casual, when I need it, without having to
justify it.  For most professionals, nobody is looking to see how many
compute cycles you are wasting with your desktop PC, or the laptop I'm
typing this on.  It's just a resource that is available to you, and
unaccountable in terms of usage.

>