[Beowulf] Contents of Compute Nodes Images vs. Login Node Images

[Greg Lindahl]

On Tue, Oct 23, 2018 at 05:48:00PM +0000, Ryan Novosielski wrote:

> We’re getting some complaints that there’s not enough stuff in the
> compute node images, and that we should just boot compute nodes to
> the login node image

It's probably worth your while sitting down with your users and
learning how they want to use the tool, instead of telling them.

As an example of a usage that needs a bigger node image, the PathScale
compilers shipped with a utility named 'pathopt' that compiled your
program with 280 different sets of compiler switches, so you could
discover which ones worked great for your code on your data. This
would take quite a while to run on a single machine, but runs pretty
quickly on a cluster... if you can compile on the compute nodes.

So yeah, it's an embarrassingly parallel workload. But if it saves you
10% on a long-running job, it's worth it.

That's probably not the example your users have in mind, but there's
plenty of pre-processing and post-processing computations in many
scientific workloads.

-- greg