[Noisebridge-discuss] DL380s

[Josh Myer]

On Sat, Jul 11, 2009 at 07:37:40PM -0700, Dr. Jesus wrote:

> The machine learning people wanted to use the DL380s, but they
> couldn't keep them on in the space for noise and power reasons.

Some of the machine learning people wanted to use them; the overall
group hasn't had need of it yet.  We have lots more processing power
in the laptops sitting on the table, more than we need for what we do
on a weekly basis.  It would be nice to have massive quantities of CPU
available for 5 wall-minute reductions that chew up 20 CPU-hours, but
only once we get to certain algorithms.  It makes parameter tuning go
a hell of a lot faster, which is really handy when you're demo'ing.

FWIW, the one time ML has really needed a machine, I just spun up an
EC2 instance and handed out a login.  It worked pretty well, but gets
expensive for long-running jobs.  That said, if we can't beat the cost
of EMR, though, we might as well just have individuals (like myself)
underwrite the cost of those services.  Which is what I've been doing
in the past, and it seems to work out well enough.

(Don't read this as discouragement, just clarification.  I'd love us
to have a 150-node cluster at the ready, but it's not something we
really need for the ML group at the moment.  And, if we do need lots
of CPU, we can probably go EC2 or EMR.)
-- 
Josh Myer   650.248.3796
  josh at joshisanerd.com