[Beowulf] Roadrunner shutdown

Max R. Dechantsreiter max at performancejones.com
Fri Apr 5 13:03:06 PDT 2013


John,

> Any further talk of NOT throwing money at HPC experts is not to be tolerated.
> Indeed, I can cope with oodles of money being thrown at me and I can solve lots of problems!
>
> Seriously though - if skills like ours are that rare, and can save lots of computer resources,
> why in a capitalist (running dog) society are we not being rewarded with millions of $$$ ??

Hear, hear!

Playing back-of-the-envelope myself, it's generally fairly
easy to speed up a floating-point application by 30%, given
a few days or weeks (that has been my routine experience).
Hence, given a modestly-sized 10,000-core cluster, that
investment of time/expertise would be worth an additional
3,000 cores or thereabouts.  Not bad, eh?

Plus if scaling issues are involved, the reward could be
far greater - and a user might be able to run on a decent
fraction of the cluster's cores.

In the case of Roadrunner, taking it on faith that the
application was neutron transport, single-threaded tuning
and algorithm mapping would almost surely have bee primary,
for Monte Carlo simulations are almost all embarrassingly
parallel....

Unfortunately vendors rarely assist users after the sale,
and administrators would rather have more hardware with
concomitant rise in Top500 ranking.  More's the pity!

Regards,

Max



More information about the Beowulf mailing list