[New Standard]: Clarify bit-for-bit reproducibility requirement and consider relaxing this in operations.

[JacobCarley-NOAA]

What new standard is being proposed?

At present there is no mention of a bit-for-bit reproducibility requirement in the current standards. In fact, this is only noted in the EE2 process document. I would like to propose two items:

1. We clarify the requirement here in a way that supersedes the one noted in the EE2 document.
2. We consider relaxing the requirement for bit-for-bit reproducible restarts in operations.

Why is the new standard necessary?

1. The EE2 document is not the correct place for requirements. This should be in the standards document.
2. Requiring bit-for-bit restarts in operations is excessive and eliminates all possibility of using compression algorithms that would help speed up I/O and reduce file size (easing disk burden and reducing the bandwidth strain on our networks). We should consider making it clear that bit-for-bit restarts are not required.

How will the new standard be enforced?

I don't think there is anything to enforce here. It is more or less to add clarity in the standard and relax the requirement that is described in EE2.

Impact of this standard?

Pros: Faster I/O, lower disk footprints, less bandwidth usage.
Cons: No longer bit reproducible on restart in operations.

One thing to consider: we should still enforce this requirement at the repository level. I think it's still important that codes reproduce for scientific integrity.

[ligiabernardet]

One thing to consider: we should still enforce this requirement at the repository level. I think it's still important that codes reproduce for scientific integrity.

This makes sense to me. It would be good to discuss whether reproducibility is expected/needed is processor layouts change for the restart. It is my understanding that some codes do reproduce under these layout changes, while others do not.