[cfarm-users] Is it fine to run a micro-benchmark on multiple cfarm machines?

[Zach van Rijn]

On Wed, 2026-10-07 at 14:38 +0000, Matthew Malcomson via cfarm-
users wrote:
> Hello,
> 
> I'm interested in knowing how long the existing default GNU
> OpenMP spin-wait takes on various different machines.
> I'd like to use the cfarm machines for running this experiment
> — a micro-benchmark that should take a few seconds, but run on
> all machines I have access to.
> 
> My questions are:
>    1. Is there anything I need to watch out for some machines
> I shouldn't access, access patterns that are more friendly,
> etc?

Ensure that each machine is either reasonably idle or that you
won't be risking interfering with anyone else's experiments.

Benchmarks are fine without notice if they're short-lived; if
you need to run long experiments then we could schedule them.

I would draw a line (in general) at network-related benchmarks
or anything that would potentially cause disruptions for the
host of a given machine. That wouldn't apply here, but just FYI.

Another thing to watch for is generating large amounts of data
and filling up the limited disk space that some machines have.

Anything likely to cause DoS or data loss would be a problem.


>    2. Is this general approach a fine use of the cfarm
> machines?  (I've not used them before).

Yes. This is a good application for CFarm resources.

> 
> Context:
> I'm working on libgomp and I'd like to gather evidence that
> the default spincount no longer estimates the 3msec that it
> was written to target (I'd like to increase the spincount,
> Jakub seems to think feasible if I gather the right evidence).
> I can see this on the machines I have access to, but would
> like more machines in the data set.

Seems reasonable to me. Thanks for asking.


Z