RAxML - 8.2.12

Basic information

Installation

This entry covers the entire process performed for the installation and configuration of RAxML on a cluster with the conditions described below.

Usage

This subsection describes how to use RAxML on a cluster and the necessary elements to get a good performance.

  1. Before launch RAxML you should read next documentation

    Note

    It is really important to understand how the HYBRID version works, since this is the only available version for HPC scenarios. Additionally, understanding the behavior of the HYBRID version is the key to properly use the computational resources and achieve better performance.

  2. In the following example we will run 100 bootstrap replicates (MPI parallelization) and independent tree searches (PThreads - shared memory) for each bootstrap replicate, all of this using SLURM (Resource Manager) to spawn properly the processes across the nodes.

    slurm.sh
    #!/bin/bash

#SBATCH --partition=longjobs
#SBATCH --nodes=3
#SBATCH --ntasks-per-node=2
#SBATCH --cpus-per-task=16
#SBATCH --time=48:00:00
#SBATCH --job-name=RAxML_test
#SBATCH -o result_%N_%j.out
#SBATCH -e result_%N_%j.err

# Default variables
export SBATCH_EXPORT=NONE
export OMP_NUM_THREADS=1

# Load RAxML module file
module load raxml/8.2.12_intel-17.0.1

# Launch RAxML (MPI with srun (pmi2) and PThreads using '-T' argument 
# and SLURM_CPUS_PER_TASK environment variable.
srun --mpi=pmi2 raxmlHPC-HYBRID-AVX2 -s funiji.fasta -T $SLURM_CPUS_PER_TASK \
     -X -f a -n out_1 -m GTRGAMMA -x 6669 -p 2539 -N 100

    Note

    Node quick specs (Apolo II): 32 Cores, 64 GB RAM

    • - -ntasks-per-node → MPI process per node
    • - -cpus-per-task → PThreads per MPI process
    • - -nodes → Number of nodes

    In this case, we will use 2 MPI process per node and each MPI process has 16 PThreads; for a total of 32 processes per node. Also we will use 3 nodes.

Authors