was OT -> Re: [Beowulf] SuSE 9.3

Cristian Tibirna ctibirna at giref.ulaval.ca
Wed Jul 13 12:50:46 PDT 2005

On 13 July 2005 13:29, Jim Lux wrote:
> Aside from compatibility issues, there's a not-insignificant cost to
> rolling out a change to thousands of desktops, some small fraction of which
> WILL break for one reason or another, no matter what OS you're
> running.  Say you're a corporate IT manager responsible for 10,000 desktop
> machines.  If 0.1% of those machines break (which is a very small number),
> you've got to handle 10 calls, each of which probably costs you somewhere
> between $500-1000 (staff to handle it, the lost productivity of the person
> who's machine broke, etc.).
> Say you roll out the change in the daytime.. there's going to be some
> disruption of what's going on with each desktop.  Say it costs 15 minutes
> for each user.. times 10,000 users, that's 2500 work hours, conservatively
> well over $100K worth.  Assuming all goes well. If some fraction of those
> users decide to call the help line because "something weird is going on
> with my PC", you've just radically increased the cost of the roll out.
> So, you say, roll it out at night.  Then, some fairly significant fraction
> of the machines won't get the update because the user has turned it off
> (despite broadcast messages and exhortations to "please leave your computer
> on tonight").
> This is more a manifestation of having thousands of machines in the hands
> of unsophisticated users, than any particular OS choice. By the way,
> sophisticated users are actually worse: They notice that something weird is
> going on and call to ask; They're more likely to have changed the "default
> configuration" of the system; They're more likely to have installed some
> other software, outside the official configuration management regime.
> So, moderate to big shops tend to want to avoid willy-nilly rearrangement
> of the computing configuration.  Once a year is nice.. You budget $500K or
> so for the rollout and its costs (pretesting, support, organization, etc.)
> and you're done with it.

So long live braindead terminals connected in X, VNC, NX (especially the 
later) to the _very_ few central servers, which can be -- yes, let's come 
back on topic -- Linux clusters, even beowulf ones, even running the other 
damn os in virtual machines; clusters equipped with HA (APM, backup, 
whatever). And all these costs you mention will simply go away. _And_ you 
gain the freedom of having users run roaming sessions, even from their car in 
the parking lot outside, if needed, as well as the freedom of running huge 
jobs on readily available numerical behemots if ever need to (granted that 
all code would finally be written with parallelism in mind).

It's a bitter irony that we come after 35 years (yeah, almost as long as I 
actually lived) back to the realization that, at least with IT, some amount 
of centralization is better. Then, upgrades don't depend on freaks or dumbs 
or on factor 10e5 plannings.

Cristian Tibirna				(418) 656-2131 / 4340
  Laval University - Québec, CAN ... http://www.giref.ulaval.ca/~ctibirna
  Research professional - GIREF ... ctibirna at giref.ulaval.ca

More information about the Beowulf mailing list