[Beowulf] Re: MS Cray

[Robert G. Brown]

On Thu, 18 Sep 2008, Lux, James P wrote:

> I think the model that John Vert mentioned, using it as a software
> development workstation to try things out before running on the "big
> iron" is actually probably a more likely scenario. And for that, you
> might not want the full up configuration, just enough to make it a
> "real" cluster so you can work out the interprocessor communications
> issues.

Which is fine, if the "big iron" (more likely "big bad wulf":-) you want
is running the same interconnect, etc.  In which case as few as two
blades might work, but then, so would two towers at about the same
footprint (in all dimensions) and half the cost.

A bladed system full, well, that makes sense of a sort as it is the only
way to get that density, sort of, except as Mark pointed out the
advantage isn't quite as pronounced as perhaps it once was.  A bladed
system with two or three blades and five or six empty slots?  That's
just plain dumb.  Even my desk, swollen as it is with the flood of
papers I'll get around looking at the day Satan goes snow skiing down
the slopes of hell, can be trivially rearranged to add two more
mid-towers to the one that is already there and a small network
interface.  In fact, I've got three mid-tower chassis (two of them dead)
sitting in my office anyway.  I could probably build myself a 17 core
development cluster out of spare parts and a couple of new
motherboard/CPU/memory sets for $4K, or spend even less on just single
quad motherboards and a better network for 9 cores.

As is all too often true, a niche product has narrow boundaries on the
places it SANELY makes sense, although naturally that is almost
irrelevant to the issue of marketing it.  And hey, Penguin's "blade
center" is awfully pretty and awfully high density if one ever could
afford to buy it fully populated.

One of my earliest personal cases of optimism was buying an $80K SGI
220S (think of it as a "very early bladed server" the size of a
refrigerator and drawing anywhere from 1-3 KW) with just one dual CPU
blade, thinking we'd eventually get more at $20-30K/blade.  Ha ha ha.
We sold it six or so years later for $3000, when we could get over the
counter Sparcstations that were STILL expensive at $5K or thereabouts
but which were even faster.

But people do buy them -- grab a copy of PC Mag or Linux Mag (now
sorta-defunct?) or whatever and leaf it.  What do you find?  Lots of ads
for rackmounts.  A few ads for consumer systems (but not that many
anymore, outside of laptops).  And yeah, quite a few ads for
bladeservers.

Just not USUALLY for cluster computing, I don't think, at least by
people on this list, which is a last bastion for non-turnkey beowulfery
where we always ask ourselves if it is better to buy a prettier and
niftier system or get more processors, and almost invariably select
"more processors" in an uglier form factor when we find it.

:-)

    rgb

> Jim
>

-- 
Robert G. Brown                            Phone(cell): 1-919-280-8443
Duke University Physics Dept, Box 90305
Durham, N.C. 27708-0305
Web: http://www.phy.duke.edu/~rgb
Book of Lilith Website: http://www.phy.duke.edu/~rgb/Lilith/Lilith.php
Lulu Bookstore: http://stores.lulu.com/store.php?fAcctID=877977