[oclug] [OT] supercomputers

[Brad Barnett]

On Sun, 16 Nov 2003 13:08:11 -0500 (EST)
Jon Earle <je_oclug at kronos.honk.org> wrote:

> On Sun, 16 Nov 2003, Matthew Wilcox wrote:
> 
> > The thing about clusters is that you start to *really* care about
> > MTBF. If you have a MTBF of 10 years and 2000 units, you're replacing
> > a unit every 2 days.  Clusters are great, but they're not a panacea. 
> > You also care about power consumption and aircon ;-)
> 
> I suspect though, that the MTBF for PC units is lower, likely closer to
> 5yrs.  With 200 PCs in a cluster (still a large cluster), that comes to
> about 91 days between unit replacements.  (A 10yr MTBF would result in
> replacement of a unit every 6 months.)
> 
>

One thing to keep in mind is that a server PC and a workstation / home
user PC are two different things.

You can most definitely buy high grade motherboards, cases, ram, you name
it.  These are still significantly cheaper, by a long shot, than our
friendly "big box", and offer a much longer lifespan than standard, off
the shelf components.  You can easily slap a nice hot-swappable hardware
raid, dual power supply and so on in as well, whilst keeping costs still
very minimal.

Sure, it won't eliminate the problem discussed, but it will reduce it by
extending the time to failure on average.