Exploring parallel MPI fault tolerance mechanisms for phylogenetic inference with RAxML-NG

Area of Science:

• Computational Biology
• Bioinformatics

Background:

• Phylogenetic tree inference is increasingly performed on large-scale computing systems.
• The parallel scalability of these tools has improved, but fault tolerance remains a challenge.
• Handling molecular data requires robust computational methods.

Purpose of the Study:

• To explore parallel fault tolerance mechanisms for phylogenetic inference tools.
• To assess software modifications and performance penalties of enabling fault tolerance in RAxML-NG.
• To ensure reliable phylogenetic tree inference on high-performance computing systems.

Main Methods:

• Implemented parallel fault tolerance mechanisms and algorithms in RAxML-NG.
• Modified RAxML-NG software to incorporate recovery mechanisms.
• Evaluated performance penalties and conducted failure simulations.

Main Results:

• The slowdown from recovery mechanisms in RAxML-NG averaged 1.00 ± 0.04.
• Overall slowdown with a fault-tolerant Message Passing Interface was 1.7 ± 0.6 for large datasets.
• RAxML-NG successfully recovered from multiple, subsequent, and checkpointing failures automatically.

Conclusions:

• Parallel fault tolerance can be effectively implemented in phylogenetic inference tools like RAxML-NG.
• The performance overhead is acceptable, ensuring reliable large-scale phylogenetic analyses.
• The fault-tolerant RAxML-NG code is publicly available for use.