gromacs-md¶

GPU-accelerated molecular dynamics. GROMACS with hydrogen-mass repartitioning (4 fs timestep), exposed as an async HTTP service.

Pre-reqs¶

NVIDIA GPU (L4/A10G at minimum for reasonable throughput; L40S/A100/H100 preferred)
Docker + NVIDIA Container Toolkit
CUDA 12.x drivers on the host
~8 GB GPU memory per concurrent simulation

Deploy¶

Via docker-compose (default):

Uncomment the gromacs-md block in docker-compose.yml, then:

docker compose up gromacs-md

Standalone (remote GPU box):

docker run -d \
  --name novomcp-gromacs-md \
  --gpus all \
  -p 8024:8024 \
  -e API_KEY=<choose-a-key> \
  -v /path/to/persistent/results:/results \
  --restart unless-stopped \
  ghcr.io/novomcp/gromacs-md:latest

Wire into the engine¶

export GROMACS_MD_URL=http://localhost:8024
# or remote:
export GROMACS_MD_URL=http://gpu-box.local:8024
export GROMACS_API_KEY=<the same key>   # sent as the API-Key header

Verify¶

curl -s http://localhost:8024/health
# {"status":"healthy","gpu_available":true,"gromacs_version":"2023.5"}

Tools that light up¶

run_molecular_dynamics, async MD (poll job status via get_job_status)
generate_dynamics, same, higher-level wrapper with defaults

Async pattern¶

MD jobs are long-running (minutes to hours). The engine returns a job_id immediately; poll status:

curl -s http://localhost:8018/mcp/tools/get_job_status \
  -H 'Authorization: Bearer x' \
  -H 'Content-Type: application/json' \
  -d '{"arguments":{"job_id":"<the-id-you-got-back>"}}'

Or set email in the original call to get notified on completion (requires Resend API key via env; see below).

Env vars¶

Var	Default	Purpose
`PORT`	`8024`	HTTP listen port
`API_KEY`	(required)	Key the engine must send; the service refuses jobs without it
`REDIS_URL`	`redis://localhost:6379`	Job status store (`get_job_status` reads it)
`AUTODOCK_GPU_URL`	(unset)	autodock-gpu endpoint. Protein-ligand runs dock the ligand first and start MD from the top pose; unset, the ligand is centred on the co-crystallized ligand's site (result field `ligand_placement`)
`MDRUN_GPU_RETRIES`	`2`	Retries of a failed GPU `mdrun` stage on the GPU
`MDRUN_CPU_FALLBACK`	`1`	After those, rerun the stage once on the CPU (result field `mdrun_retries`); `0` fails the job
`MAX_CONCURRENT`	`1`	Simulations per GPU
`RESULTS_DIR`	`/results`	Where trajectories land
`RESEND_API_KEY`	(unset)	Email notifications on job completion
`HMR_HYDROGEN_MASS_DA`	`3.0`	HMR hydrogen mass (3 Da → 4 fs timestep)

Speed + cost¶

10 ns simulation on dasatinib•ABL kinase (~55K atoms): ~18 minutes end-to-end (solvate + parameterize + minimize + equilibrate + 10 ns) on L40S. Sustained rate >1 µs/day.
Cheapest cloud spot: g5.xlarge (A10G) at $0.30/hr spot; usable for short runs and hackathon-scale exploration.
For multi-µs production simulations, prefer A100 40GB or H100.