[Chicago] What is cloud computing good for?

Garrett Smith g at rre.tt
Mon Aug 9 21:26:07 CEST 2010


On Mon, Aug 9, 2010 at 1:58 PM,  <skip at pobox.com> wrote:
>
>    Massimo> If you are not interested in web apps (and looks like you are
>    Massimo> not) you may be interested in this:
>
>    Massimo> http://www.picloud.com/
>
> You are correct.  I care not one whit about web apps.  I'm interested solely
> in single-threaded compute-intensive tasks.  Currently, we run the Torque
> batch queuing system on Solaris.  It works fine, though I am constantly
> having to scramble to find hardware.  I might have 10 cores available during
> the day, 50 at night.  That works well until I'm asked to run 4000 jobs one
> day.  Each job can run from a few minutes to several hours.
>
> PiCloud looks interesting except it doesn't seem that your Python functions
> can use C or C++ code which is not part of Python proper.  While I write
> almost all my software in Python I rely heavily on internal libraries
> written in C++ by other groups here at work.

This is pretty recent from Amazon...

http://aws.amazon.com/elasticmapreduce/

I'm not actually the Amazon fanboy that I appear to be. But I
impressed with how they've built AWS out over the last few years.

For my own understanding, I parse "cloud" as "no physical
infrastructure". This actually is a big deal, IMO -- not because it's
new (the mainframe business was largely one of leasing compute cycles
and data storage) -- but because fixed costs of using these resources
are getting very close to $0.

Jason's absolutely right -- at a certain point it's more cost
effective to invest in your own infrastructure. Cloud offerings are
generally a little bit cheaper than dedicated hardware but are also *a
lot* less performant. But I've found that companies are willing to put
up with this cost equation for the convenience of changing (or walking
away from) their operations at any time.

I generally recommend:

- Develop and test in a virtualized (i.e. cloud) environment
- When the economics over a multi-year period make sense, convert to a
"real" infrastructure - this is simply a price/performance decision

The cloud providers that can provide a combination of virtual and
real/dedicated hardware -- that is, can simplify the migration from
dev/test/early-production to long term production -- will have a super
compelling value proposition.

Garrett


More information about the Chicago mailing list