[clug] Interesting article

Andrew Janke a.janke at gmail.com
Mon Jan 5 11:56:43 GMT 2009


> Commodity gear is cheap. Managing said gear is not,

Can we modify that last posit to "Managing said gear is usually not"? :)

> It's much easier (cheaper) to manage fewer powerful
> blade servers or massive virtual servers than piles of rinky-dink
> commodity gear.

"Perhaps". In my experience it is not really that much different given
what said "good stuff" costs when it breaks. For example I just had a
BBU die in a largeish SATAII RAID controller last week, replacement
cost is about $400 or so...  If I was running a commodity version
using gfs or the likes across a network, I could have bought a whole
extra TB with those $$.

The other thing I have always liked about piles of heterogeneous gear
is that the "upgrade path" has very little resistance, You just keep
buying a few new faster ones all the time and slowly let the old ones
drop of the bottom. I spent a few years in Montreal at a research lab
that supported ~200 biomed engineering types doing fairly intensive
computational neuroimaging. Meaning lots of batch jobs. My job was to
formulate a plan to upgrade their existing $6million investment in the
"supercomputing" of the time. (SGI 3800 series origin gear). Of course
in the end this meant shifting the main processing to the existing
desktop machines with a bit of liberal application of cfengine +
Debian.

The plan was to "limp" along like this until more funding was found to
replace the existing SGI monster. These plans got rapidly shelved when
it became rapidly apparent that the desktop machines were outstripping
the 160CPU SGI monster 2:1 despite all the harbingers of doom
predicting NFS lockups and the likes. The eventual plan also involved
distributed redundant 4U machines with 6TB of storage in them each
instead of central file servers.

This system is still used to this day there (www2.bic.mni.mcgill.ca)
as the idea of just keeping on slowly updating desktop machines and
using the spare CPU cycles just works (TM). It is all supported by 2
sysadmins for the 150+ debian/Ubuntu desktops there.

What I _can_ guarantee you though is that if you go for blade servers
and the likes, it will cost a heap (including the maintenance
contracts) and they will have to be junked in a few years as they will
be too expensive to upgrade. Despite what the rep told you regarding
"upgrade paths".

Of course YMMV.


-- 
Andrew Janke
(a.janke at gmail.com || http://a.janke.googlepages.com/)
Canberra->Australia    +61 (402) 700 883


More information about the linux mailing list