MetaHipMer2 (MHM2)

MetaHipMer2 is a de novo metagenome short-read assembler written in UPC++ and CUDA. It runs efficiently on single servers and on multinode supercomputers, where it can scale up to coassemble terabase-sized metagenomes. On Roihu it is provided as a GPU-accelerated build for the GH200 GPU nodes.

MetaHipMer2 (MHM2)

License

MetaHipMer2 is free and open source under a Berkeley Lab modified BSD 3-Clause license.

Available

Roihu-GPU: 2.2.2.0-20260904, via the bio-apps module (GPU nodes only).

MHM2 runs on the Roihu GPU (GH200) nodes. Log in to roihu-gpu.csc.fi and load the modules there so you get the GPU build.

Usage

MHM2 is part of the bio-apps collection on Roihu. Load the bio-apps module tree and then the MHM2 module:

module load bio-apps/v202603
module load mhm2/2.2.2.0-20260904

Assemblies are launched with the mhm2.py driver, which starts the parallel (UPC++) processes for you — do not prefix it with srun. A basic run on paired-end reads:

mhm2.py -p reads_1.fastq,reads_2.fastq -o assembly_result

Use -p/--paired-reads for paired reads split into separate R1/R2 files (the two files are comma-separated), or -r/--reads for interleaved paired reads in a single file.

The assembled contigs are written to assembly_result/final_assembly.fasta. For the full list of options use:

mhm2.py -h

Assembling metagenomic data is resource demanding, so run MHM2 as a batch job on a GPU node.

MHM2 requires a GPU node

MHM2 is a GPU build and links directly against the CUDA driver library, so it runs only on Roihu GPU nodes. On a login node it fails at startup with error while loading shared libraries: libcuda.so.1 because login nodes have no GPUs.

Example batch script

MHM2 is process-parallel (UPC++): mhm2.py launches one process per Slurm task and derives that count from --ntasks-per-node. Reserve one task per CPU core you want to use, with --cpus-per-task=1. The example below uses one GPU with 72 processes:

#!/bin/bash
#SBATCH --job-name=mhm2
#SBATCH --account=<project>
#SBATCH --output=mhm2_out_%j.txt
#SBATCH --error=mhm2_err_%j.txt
#SBATCH --partition=gpumedium
#SBATCH --time=12:00:00
#SBATCH --nodes=1
#SBATCH --ntasks-per-node=72 --cpus-per-task=1
#SBATCH --gres=gpu:gh200:1

module load bio-apps/v202603
module load mhm2/2.2.2.0-20260904

mhm2.py -p reads_1.fastq,reads_2.fastq -o assembly_result

Replace <project> with your CSC project (for example project_2001234). You can use csc-projects to check your CSC projects. Submit the job with sbatch:

sbatch mhm2_job.sh

If you reserve only one task (the default), MHM2 runs on a single core and warns Detected slurm job restricts cores to 1 because of SLURM_TASKS_PER_NODE=1 — set --ntasks-per-node as shown to avoid this. To use a full GPU node, request all four GPUs and 288 tasks (--ntasks-per-node=288 --gres=gpu:gh200:4); MHM2 shares the node's GPUs among its processes. Because it communicates over InfiniBand it can also coassemble across several nodes — increase --nodes accordingly, and also switch to the gpularge partition which accommadates requests for multiple nodes.

Match the number of processes to the data size

MHM2's parallel start-up, communication and GPU-sharing overheads dominate on small inputs, so many processes can be slower than a few. For small or test datasets, request only a handful of cores (for example --ntasks-per-node=4 --cpus-per-task=1 on the short-lived gputest partition). Use the full 72 cores per GPU, or a whole node, only for large metagenomes.

See creating a batch job script for Roihu for more about batch jobs, and the GPU partitions for the available GPU resources.

Support

CSC Service Desk

More information

MetaHipMer2 repository