メニュー
MIT Engaging cluster reference. SLURM job scheduling, Coley group nodes, filesystem layout, Python environments, and common commands. Use when working with the Engaging HPC cluster.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
| name | engaging |
| description | MIT Engaging cluster reference. SLURM job scheduling, Coley group nodes, filesystem layout, Python environments, and common commands. Use when working with the Engaging HPC cluster. |
SLURM-managed HPC cluster run by MIT ORCD. Kerberos + Duo 2FA required (reducible via ORCD OnDemand or SSH control channels).
Tips adapted from the Coley Group Engaging Guide by Sam Goldman, David Graff, Runzhong Wang, and Magdalena Lederbauer.
Login node (Rocky 8):
ssh username@orcd-login.mit.edu
Two-Factor Authentication (Kerberos password + Duo) required. Reduce repeated 2FA by:
SSH config for convenience:
Host engaging
Hostname orcd-login.mit.edu
User <kerberos>
ControlMaster auto
ControlPath ~/.ssh/control-%r@%h:%p
ControlPersist 8h
| Partition | Node | CPUs | GPUs | GPU VRAM | RAM |
|---|---|---|---|---|---|
| sched_mit_ccoley | node1236 | 52 (2x Intel 6230R) | 8x 2080 Ti | 11 GB | 514 GB |
| sched_mit_ccoley | node1238 | 128 (2x AMD EPYC 7702) | — | — | 515 GB |
| pi_ccoley | node2436 | 64 (2x Intel 8562Y+) | 4x H100 | 80 GB | 1000 GB |
| pi_ccoley | node2519-2522 | 64 (2x AMD EPYC 7513) | — | — | 249 GB |
Resource etiquette for node1236: 52 CPUs and 514 GB across 8 GPUs. Use -n 6 --mem-per-cpu=8G per single-GPU job so all 8 GPUs can run simultaneously. Requesting too many CPUs or too much memory per job blocks other GPUs from being used.
Prefer CPU-only nodes (node1238, node2519-2522) for non-GPU work.
| Partition | Node | CPUs | GPUs | GPU VRAM | RAM | Notes |
|---|---|---|---|---|---|---|
| ou_cheme | node4110-4113, 4115-4116 | 128 (2x64 cores) | — | — | 770 GB | also in mit_preemptable |
| ou_cheme | node4114, 4117 | 128 (2x64 cores) | — | — | 1540 GB | high-memory, also in mit_preemptable |
| ou_cheme_gpu | node5000-5001 | 120 (2x60 cores) | 8x H200 | 141 GB | 2055 GB | also in mit_preemptable |
| ou_cheme_gpu | node5003-5005 | 128 (2x64 cores) | 8x RTX PRO 6000 | 96 GB | 2055 GB | also in mit_preemptable |
All ou_cheme CPU nodes are in both ou_cheme and ou_cheme_preemptable partitions, and also in mit_preemptable. GPU nodes are in ou_cheme_gpu and mit_preemptable.
| Partition | Purpose | GPUs | Max Time |
|---|---|---|---|
| mit_normal | CPU batch/interactive | — | 12 hrs |
| mit_normal_gpu | GPU jobs | L40S (44 GB), H100 (80 GB), H200 (140 GB) | 6 hrs |
| mit_quicktest | Quick testing | — | 15 min |
| mit_preemptable | Low-priority, may be preempted | A100/L40S/H100/H200 | 48 hrs |
Request GPUs with -G [type]:[count] (e.g., -G l40s:1, -G h200:2). Default on mit_normal_gpu is L40S.
CPUs per GPU on mit_normal_gpu: L40S = 16, H200 = 15. Requesting more CPUs than this wastes a GPU slot.
| Path | Size | Backed Up | Purpose |
|---|---|---|---|
/home/<user> | 200 GB | Yes (snapshots) | Code, config, small files |
/home/<user>/orcd/pool | 1 TB | No | Larger datasets (not I/O intensive) |
/home/<user>/orcd/scratch | 1 TB | No | I/O-heavy job data (flash storage) |
/nfs/ccoleylab001 | 10 TB shared | No | Group shared data — don't be greedy |
/nobackup | cluster-global | No | Semi-persistent intermediate files (periodically wiped) |
/tmp | node-local | No | Fastest I/O, but not shared between nodes; gone after job ends |
Storage hierarchy for jobs: Use /tmp or scratch for active computation. Stage data from pool. Keep code and configs in home.
Scratch files deleted after 6 months of inactivity. /nobackup wiped ~monthly.
# check partition availability
sinfo -p sched_mit_ccoley
sinfo -p mit_normal_gpu -O Partition,Nodes,CPUs,Memory,Gres
# interactive session (1 GPU on group node)
salloc -p pi_ccoley --gres=gpu:1 -n 6 --mem-per-cpu=8G --time=3:00:00
# interactive session (public H100)
salloc -p mit_normal_gpu -G h100:1 -c 4 --mem=32G --time=6:00:00
# batch job
sbatch myjob.sh
# check your jobs
squeue --me
# cancel a job
scancel <jobid>
# job history and memory usage
sacct -j <jobid> -o JobID,JobName,State,ReqMem,MaxRSS,Elapsed --units=G
# check node resources
scontrol show node node1236
# who's using our partition
squeue -p sched_mit_ccoley
#!/bin/bash
#SBATCH -p pi_ccoley
#SBATCH --gres=gpu:1
#SBATCH -n 6
#SBATCH --mem-per-cpu=8G
#SBATCH -t 0-08:00:00
#SBATCH -J my_job
#SBATCH --output=logs/%j_%x.log
module load miniforge
source activate myenv
python train.py
#!/bin/bash
#SBATCH -p pi_ccoley
#SBATCH -n 6
#SBATCH --mem-per-cpu=8G
#SBATCH -t 0-05:00:00
#SBATCH --output=logs/%j_%x.log
module load miniforge
source activate myenv
eval $CMD
Launch with: sbatch --export=CMD="python train.py --lr 1e-4" generic_slurm.sh
Use --requeue flag so jobs restart when preempted. Ensure your code checkpoints regularly.
sbatch --requeue -p mit_preemptable -G l40s:1 myjob.sh
# load system miniforge (do NOT install your own conda/miniconda)
module load miniforge
# option 1: conda/mamba environment
mamba create -n myenv python=3.12 pytorch torchvision
source activate myenv
# install pip-only packages after activating:
pip install some_package # do NOT use --user flag
# option 2: virtual environment
python -m venv ~/myproject/venv
source ~/myproject/venv/bin/activate
pip install -r requirements.txt # do NOT use --user flag
In job scripts, always load miniforge and activate the environment explicitly — don't rely on login shell state:
module load miniforge
source activate myenv
conda init or mamba init — it modifies .bashrc and causes issuespip install --user — installs to ~/.local instead of the environmentPYTHONPATH unless absolutely necessaryMove them to scratch by creating ~/.condarc:
envs_dirs:
- /home/USERNAME/orcd/scratch/.conda/envs
pkgs_dirs:
- /home/USERNAME/orcd/scratch/.conda/pkgs
auto_activate_base: false
Back up environments with conda export --name myenv --file=environment.yaml.
Port-forwarding approach (better GPU than OOD's default K20):
# terminal 1: get a GPU allocation
ssh orcd-login.mit.edu
salloc -p sched_mit_ccoley -w node1236 --gres=gpu:1 -n 6 --mem-per-cpu=8G --time=3:00:00
# note the allocated node, e.g., node1236
# terminal 2: port forward through login node to compute node
ssh orcd-login.mit.edu -L 9998:localhost:9998
ssh node1236 -L 9998:localhost:9998
jupyter lab --port 9998
# from local machine to engaging
scp -r localdir/ username@orcd-login.mit.edu:~/project/
# or with rsync (better for repeated syncs)
rsync -az --exclude '.venv*' --exclude '__pycache__' --exclude '.git' \
. username@orcd-login.mit.edu:~/project/
which python points to your env. Load miniforge + activate env in the job script, not just at login.eval "$(conda shell.bash hook)" before conda activate.du -sh ~/.conda), move to scratch (see above). Remove unused envs with conda remove -n old_env --all.sinfo for node states.which python to verify. Virtual envs inherit the python used to create them.Verifies that bibliography references exist and have accurate metadata. Use when the user asks to "check references", "verify a bibliography", "validate citations", "check for hallucinated citations", "check a .bib file", or mentions auditing a BibTeX file for accuracy. Looks up each entry against Crossref and arXiv, then spawns web-search subagents for non-indexed sources, flagging missing DOIs, wrong years, title mismatches, venue mismatches, and broken URLs.
Reference guide for the MolGPU shared GPU cluster. Machine specs, SSH setup, shared filesystem, conda/mamba, multi-GPU training tips, and common commands.
VASP on MIT Engaging. VASP 6.4.2 (GCC/OpenMPI, Rocky 8 compatible), module loading, SLURM submission, POTCAR paths. Use when setting up or running VASP calculations on Engaging.
MassSpecGym benchmark reference — dataset schema, Python API, transforms, evaluation metrics, retrieval pipeline, and common patterns. Use when working with MassSpecGym data, models, or evaluation.
Machine learning workflows for molecular data. Use when building ML models with molecules, splitting datasets, selecting fingerprints for ML, avoiding data leakage, or doing scaffold-based train/test splits. Covers common pitfalls in molecular ML.
Python package and environment management using uv and mamba. Use when installing packages, creating virtual environments, setting up new projects, or managing dependencies. NOT for general Python coding questions.