Adjust process requirements to test_full case.

[charles-plessy]

Description of feature

Run the test_full set again; look at usage metrics; adjust module resources (high, low, …) accordingly.

[charles-plessy]

A benchmark on vertebrate genomes with the RY4 seed shows that ALIGNMENT_SPLIT_M2O never used more than 12 GB, 8 CPUs and 4 hours (process_low). However, it took 12–18 GB (Q1–Q3) on primates (test_full) with the RY128 seed.

With the RY128 seed, ALIGNMENT_SPLIT_O2M used the largest amount of resources for SPLIT processes, between 13 and 23 GB (Q1–Q3). Job durations were never longer than one hour.

Based on this, reducing LAST_SPLIT from process_high to process_medium (36 GB and 6 CPUs for 8h) would allow for more parallel tasks to run in most cases.

[charles-plessy]

The same benchmark (that used the RY4 seed) shows that ALIGNMENT_TRAIN and ALIGNMENT_LASTAL_M2O scale their memory consumption with the length of the target genome. Their CPU efficiency decreases with longer genomes. On pairs of long target and query genomes, ALIGNMENT_LASTAL_M2O will sometimes take more than 16h (but may fail earlier because needing more than 72 GB memory. On primate genomes with the RY128 seed, it peaked at 40 GB for a human–human comparison.

When LAST_TRAIN overuses resources, it also finishes fast, and for large genomes it really needs process_high, so let's not change it.

[charles-plessy]

Merged in bc33998