PB v4.6.0 optimisation

[fredjaya]

Parabricks v4.6.0 requires benchmarking and re-optimisation on Gadi, as gpuvolta is not compatible with this version.

Important resources:

• https://docs.nvidia.com/clara/parabricks/latest/gettingstarted/bestperformance.html
• https://github.com/kisarur/deepvariant-nextflow
• https://github.com/Sydney-Informatics-Hub/HPC_usage_reports
• https://github.com/Sydney-Informatics-Hub/Gadi-benchmarking/tree/main
• https://zenodo.org/records/6930814

Example benchmark setup

Gadi Queue	NGPUs^	NCPUS	SUs	Duration
dgxa100	2	32
dgxa100	4	64
dgxa100	8	128
gpuhopper	2	32
gpuhopper	4	64
gpuhopper	8	128

^ 1 GPU not recommended

Tuning options

• Test --gpuwrite for fq2bam throughput improvements.
• Enable DeepVariant multi-GPU mode (significant acceleration). 2
• Avoid low-memory flags (A100 has enough GPU RAM).
• Benchmark fq2bam scatter-gather (split FASTQs).
• Tune DeepVariant batch size on >4 GPUs.
• Test DeepVariant FP16 mode (Hopper‑optimised).
• Verify GPU RAM sufficiency (H100 never requires low-memory mode). 1