MiniDMR Local Cluster
MiniDMR is a CLI tool for creating and managing local Docker-based DMR clusters. It is the recommended way to run DMR locally for demos, development, and CI pipelines.
Use cases
| Use case | Description |
|---|---|
| Demos / workshops | Reproducible multi-node cluster, starts and stops cleanly |
| DMR core development | Container images with all dependencies preinstalled |
| App development | Separate image with DMR fully installed; focus on your app |
| CI pipelines | Temporary containerized cluster for integration testing |
Installation
- Linux / macOS
- Windows (PowerShell)
Latest release:
curl -fsSL https://gitlab.bsc.es/accelcom/releases/dmr/tools/minidmr/-/raw/master/scripts/install.sh | bash
Specific version (e.g. v0.0.4):
curl -fsSL https://gitlab.bsc.es/accelcom/releases/dmr/tools/minidmr/-/raw/master/scripts/install.sh | bash -s -- v0.0.4
Install to a custom directory (no sudo):
curl -fsSL .../install.sh | bash -s -- --install-dir ~/.local/bin
Latest release:
irm https://gitlab.bsc.es/accelcom/releases/dmr/tools/minidmr/-/raw/master/scripts/install.ps1 | iex
Specific version:
& ([scriptblock]::Create((irm .../install.ps1))) -Version v0.0.4
For manual installation, download the binary from the Releases page.
Commands
| Command | Description |
|---|---|
start | Start a multi-node Docker-based DMR cluster |
enter | Drop into the controller node interactively |
exec | Run an arbitrary command on the controller node |
srun | Run a Slurm job on the cluster via srun, blocking |
sbatch | Submit a batch job to the cluster via sbatch |
install | Install RPM packages with dnf in the running cluster |
upgrade | Upgrade minidmr to the latest or a specific release |
status | Show whether the cluster is running and how many nodes it has |
stop | Stop and remove all containers |
version | Print the current version |
completion | Generate shell autocompletion scripts |
help | Display help about any command |
Quick example
# Start a 4-node cluster
minidmr start --nodes 4
# Start a cluster whose image already ships with the user, injecting env vars
# into every container (controller + workers)
minidmr start --no-user-setup -e DMR_PATH=/home/malluser/dmr
# Enter the controller node
minidmr enter
# Run an arbitrary command on the controller node (e.g. to compile DMR)
minidmr exec -u malluser -- bash -c 'source env_setup.sh && ./build.sh'
# Run a Slurm job, blocking until it finishes; the exit code is propagated
minidmr srun -u malluser -e DMR_CHECKPOINT_RESTART=1 -- -N2 mpirun ./my_test
# Submit a batch job via sbatch, from a script on the controller node
minidmr sbatch -u malluser -- job.sbatch
# Or, for a quick one-off job without a script file
minidmr sbatch -u malluser -- --wrap 'hostname'
# Install packages on all cluster nodes
minidmr install blas-devel lapack-devel
# Install a package only on the controller node
minidmr install --controller-only gcc-gfortran
# Upgrade minidmr to the latest release
minidmr upgrade
# Check whether the cluster is running and how many nodes it has
minidmr status
# Same, as machine-readable JSON
minidmr status --json
# Stop and remove the cluster
minidmr stop
Running DMR tests
minidmr start --nodes 4
minidmr enter
Inside the controller node:
git clone https://gitlab.bsc.es/accelcom/releases/dmr/dmr.git
cd dmr
export DMR_PATH=$(pwd)
cd tests/ci
./dmr_full_test_run.sh compile/build_slurm4dmr_notalp.sh
The build_slurm4dmr_notalp.sh script is compatible with MiniDMR out of the box.
Global flags
| Flag | Description | Default |
|---|---|---|
--data_dir | Directory for storing cluster data and configuration | $HOME/.minihpc |
start flags
| Flag | Description | Default |
|---|---|---|
-i, --image | Container image to use | slurm-docker-cluster:slurm4dmr |
-n, --nodes | Number of slurmd nodes | 4 |
--packages-file | JSON package manifest installed after startup | $data_dir/packages.json if it exists |
--no-user-setup | Skip creating/configuring the demo user, for images that already ship with the target user | false |
-e, --env | Env var (KEY=VALUE) injected into every container; repeatable | - |
By default, start looks for a package manifest at $data_dir/packages.json and installs it when present. The manifest can be a plain array:
["blas-devel", "lapack-devel"]
Or an object with packages for all nodes and controller-only packages:
{
"packages": ["blas-devel", "lapack-devel"],
"controller": ["gcc-gfortran"]
}
install flags
| Flag | Description |
|---|---|
-c, --controller-only | Install only on the controller node |
upgrade flags
| Flag | Description |
|---|---|
-v, --version | Install a specific release tag instead of the latest |
status flags
| Flag | Description |
|---|---|
--json | Output status as JSON instead of human-readable text |
enter flags
| Flag | Description |
|---|---|
-w, --workdir | Working directory inside the container (default: mapped from current working directory) |
note
If stopped containers from a previous cluster are found (e.g. after a reboot), minidmr start resumes them instead of creating a new cluster. If --nodes is explicitly set and differs from the previous cluster size, start asks whether to remove the old cluster and create a new one.