I'm concerned that spawning an EC2 instance for each run may result in a lot
of noise in the emerged benchmarks. I've found that the hardware that
acquires each instance can differ noticeably in characteristics and shared
load from run to run, even within the same AZ. Every now and then "bad,"
underperforming, instance launches may raise spurious alarm bells. On a
large operation across multiple instances, we always expect something like a
20% difference between the first instances to report a result and the last
ones ... and are never surprised by a few failures.

The noise would be less for a single long-lived instance, or best with a
dedicated piece of hardware with no competing loads. Not that I have one I
could reliably volunteer ... but for this very specific mission, it seems
like run-to-run improvement or degradation only really has meaning if as
many other variables as we can manage are taken away. Maybe some Ruby
activist has a nice old server spinning away in their data center that feels
lonely because all the fun has gone to the cloud?

Of course, I could be fretting over nothing -- this could be in the "try it
and see" category.

On Sun, Jun 19, 2011 at 10:19 AM, Chuck Remes <[email protected]>wrote:

> I suppose I should have written that we don't *need* to benchmark every
> commit. Like the pypy project, this could be a cron job that runs once per
> day and benchmarks the latest HEAD. That would certainly eliminate the
> concern over EC2 costs.
>

-- 
--- !ruby/object:MailingList
name: rubinius-dev
view: http://groups.google.com/group/rubinius-dev?hl=en
post: [email protected]
unsubscribe: [email protected]

Reply via email to