How can I speed up runs safely?

The main controls, available in run_together and most individual commands:

• -j / --n-jobs: number of parallel jobs (pipeline-level parallelism)
• --threads: threads per job (tool-level parallelism)

The two multiply: -j 4 --threads 4 can keep about 16 cores busy.

Tuning guide

• Start with -j × --threads close to the number of CPU cores.
• Increase gradually while monitoring RAM and disk I/O.
• If the machine becomes unstable (heavy swapping, stalled I/O), reduce -j first.
• Use higher --threads when the work is dominated by a few compute-heavy steps (FICTURE training and decoding), and higher -j when there are many small steps (packaging many gene bins, many samples).

Rules of thumb

• run_together on a multi-sample run: -j parallelizes across samples and stages; per-sample packaging in a joint run also uses -j.
• Fewer gene bins (--bin-count) make packaging faster, at the cost of coarser grouping in the point layers.
• Shared servers and HPC login nodes: be conservative and follow local policy.

Example

cartloader run_together --platform 10x_xenium --in-dir IN --out-dir OUT -j 4 --threads 4

See also:

• run_together reference
• Asset Packaging reference