Do I need to edit the package setup script?

You may well have to edit the setup script depending on the target system. This includes editing compiler flags (see below for example regarding instruction sets).

Does it matter what compiler I use?

The Intel compiler collection has been used successfully for X-PSI and dependencies (namely GSL, MultiNest). We recommend first trying to use Intel in a context where performance matters.

What Intel instruction sets should I use?

If you want to test the binaries on a login node, note that you can compile with multiple instruction sets for auto-dispatch using the -x and -ax flags. See the SURFsara systems page for examples.

Model setup

Future questions and answers will be archived here.

Batch usage

Future questions and answers will be archived here.


Is I/O or disk storage a concern, or are all the files small?

I/O not a concern for likelihood calculation.

Nested sampling writes to disk at user-specified cadence (so many nested sampling iterations).

Model data such as a four-dimensional atmosphere table can be reasonably large for I/O. We recommend loading, at the outset of the run (or a resumed run), such a table into a contiguous chunk of memory for each of the Python processes running on one node. That table is pointed to for access where needed from compiled modules (C extensions to Python): it is not loaded from disk per likelihood call. We provide an example custom Python class that handles this loading (as used in Riley et al. 2019 (ApJL, 887, L21), hereafter R19).

Disk storage required is indeed small: up to \(\mathcal{O}(100)\) Mbytes for applications thus far (e.g., R19). There is a variant of MultiNest nested sampling that is much more memory and disk intensive, but we do not use it. This is because importance nested sampling is not compatible with the alternative options (read: hacks) for prior implementation (see Riley, PhD thesis).