B5 Batch independent trajectories on one GPU

Eight copies of a 64-atom cell with different seeded velocities stay eight independent simulations, not one coupled 512-atom cell. Batch.from_data_list lets one ALCHEMI run advance them together.

The cells below time one trajectory, then eight, on the same GPU. Do not run another GPU notebook meanwhile. Terminal equivalent, from the lesson checkout:

bash scripts/run-alchemi.sh --replicas 1 --cells 2 --integrator nve --warmup 10 --steps 200
bash scripts/run-alchemi.sh --replicas 8 --cells 2 --integrator nve --warmup 10 --steps 200

The notebook times each whole command, including setup and model loading. The runner’s MD-only time is a different timing window.

Show code cell source

Hide code cell source

import json
import subprocess
from time import perf_counter
from pathlib import Path
from IPython.display import Markdown, display

lesson = Path.cwd().parent if Path.cwd().name == "episodes" else Path.cwd()
single_start = perf_counter()
single_call = subprocess.run(
    ["bash", str(lesson / "scripts/run-alchemi.sh"), "--replicas", "1", "--cells", "2",
     "--integrator", "nve", "--warmup", "10", "--steps", "200"],
    check=True, capture_output=True, text=True,
)
single_wall = perf_counter() - single_start
single = json.loads(single_call.stdout.strip().splitlines()[-1])
display(Markdown(
    f"Time to finish one simulation: **{single_wall:.3f} s** "
    f"({single['measured_steps']} measured MD steps; "
    f"MD-only time {single['measured_seconds']:.3f} s)."
))

Show code cell source

Hide code cell source

batch_start = perf_counter()
batch_call = subprocess.run(
    ["bash", str(lesson / "scripts/run-alchemi.sh"), "--replicas", "8", "--cells", "2",
     "--integrator", "nve", "--warmup", "10", "--steps", "200"],
    check=True, capture_output=True, text=True,
)
batch_wall = perf_counter() - batch_start
batch = json.loads(batch_call.stdout.strip().splitlines()[-1])
display(Markdown(
    "| Simulations | Atoms each | Time to finish all (s) | MD-only time (s) |\n"
    "| ---: | ---: | ---: | ---: |\n"
    f"| 1 | {single['atoms_per_replica']} | {single_wall:.3f} | "
    f"{single['measured_seconds']:.3f} |\n"
    f"| 8 | {batch['atoms_per_replica']} | {batch_wall:.3f} | "
    f"{batch['measured_seconds']:.3f} |"
))

• The rows do different work, so their times are not a speedup ratio.

• For methods on the same eight simulations, see the B7 reviewed results; these short live runs only illustrate.

• Aggregate rate: simulations × measured steps / time to finish all (MD steps per second across all simulations, not per trajectory).

• Larger systems may already fill the GPU, so batching gains less.

Note

For NVT, change --integrator nve to langevin in both cells. ALCHEMI uses NVTLangevin; native LAMMPS can use langevin with nve/kk or a Nose–Hoover nvt/kk fix. Same target temperature does not make these interchangeable; keep NVE and NVT results in separate tables.