Open Stack

Hi Tim,

Firstly, thank-you for reading the logs and following up - it's great
to have further discussion generated!

On 2 June 2016 at 03:07, Tim Randles <trandles at lanl.gov> wrote:
> Sorry I wasn't able to make yesterday's Scientific Working Group IRC
> meeting.  The discussion looked very interesting.  I would like to +1
> Blair's second idea:
>
> <b1airo> #idea I want to run my (persistent) HPC platform/service using
> OpenStack as the infrastructure provisioning system
>
> This is exactly the use case I am exploring.  We're also in the early stages
> of defining and brainstorming support for complex work flows.

This and the complimentary facility - e.g., "datacloud" for
collaboration/sharing/post-processing/visualisation/..., close to the
traditional HPC and/or Research Data Store - seem to be the main
drivers for people I've spoken to recently that are looking to adopt
or expand OpenStack infrastructure.

> A problem that I see arising from these use case discussions is terminology.
> Let's consider the two main use cases thrown out yesterday, scientific
> clouds and cloud provisioning of HPC.  If someone wants to discuss
> "scheduling" that will likely mean wildly different things in the context of
> each use case.  Would it be useful, or even possible, to agree now on a set
> of standard definitions that we can use to avoid confusion?
>
> For instance, I propose the following:
>
> "Cloud scheduling" - provision-oriented (VMs/bare-metal) scheduling. Queued
> time may be similar to "HPC scheduling" but run time is measured in days,
> weeks, or even months. Example: Nova scheduler
>
> "HPC scheduling" - job-oriented, batch scheduling. Typically queued time and
> run time measured in hours or days. Examples: SLURM, Torque, PBS

Absolutely agree. Though I think I prefer "job scheduling" rather than
"HPC scheduling" (because HPC in itself is another overloaded term).

> I think "accounting" also falls under this overloaded-terminology rubric but
> it's much harder to define.  Even among traditional HPC centers there tends
> to be great variation in how accounting is done.

As you're probably aware, "accounting and scheduling" was one of the
activity areas prioritised by the scientific-wg meetings in Austin.
And I suspect this is one of the reasons why. Accounting needs seem to
be closely tied to the delivery model of the computing service (e.g.,
IaaS, Cluster-aaS, batch-job scheduler). Unfortunately we don't yet
have any planned deliverables or lead/s for the accounting and
scheduling activity, but perhaps a first deliverable would be to gain
a better idea of the ways in which people are approaching this now and
where their gaps are.

> What other terms have
> folks found to be confusing?

Federation is one that springs to mind.

-- 
Cheers,
~Blairo