Run the Monte Carlo π (MPI) Notebook on Perlmutter @ NERSC

This README shows how to run the provided Jupyter notebook NERSC_mpi_pi_Notebook-5.ipynb on Perlmutter, and how to execute the MPI portion on compute nodes using Slurm. This notebook demonstrates how to estimate π (pi) using the Monte Carlo "dart-throwing" method, and how to parallelize the computation using MPI (mpi4py) in a High-Performance Computing (HPC) environment.

Overview

The Monte Carlo method approximates π by randomly generating points in a square and counting how many fall inside a quarter circle. The ratio of points inside the quarter circle to the total number of points approaches π/4 as the number of points increases.

Learning Goals

• Understand the Monte Carlo method for estimating π.
• Learn basic MPI concepts: rank, size, and reduce operations.
• See how to run Python MPI programs in an HPC environment using Slurm.

---

0) Prerequisites

• Active NERSC account and an active allocation (-A <account>).
• Familiarity with Perlmutter queues: debug (short) or regular.
• The notebook: NERSC_mpi_pi_Notebook-5.ipynb is in your working directory on Perlmutter.

---

Interactive: run the notebook in NERSC Jupyter

The most convenient way to explore the notebook is through NERSC Jupyter:

1. Launch NERSC Jupyter and choose a Perlmutter CPU environment.
2. In a terminal inside Jupyter, activate your venv:

   source $SCRATCH/pm-pi-venv/bin/activate
   module load PrgEnv-gnu cray-mpich

3. Open NERSC_mpi_pi_Notebook-5.ipynb and run cells normally.
   • The serial (non-MPI) cells will run immediately.
   • For the MPI example, prefer Section 2B/3 below to execute via srun on a compute node.

Note: Jupyter kernels run as a single process; true multi-rank MPI execution should be launched with srun (below).

Run on Perlmutter with Slurm

A) Interactive test (CPU)

# Request an interactive node (CPU)
salloc -N 1 -n 8 -C cpu -q debug -t 00:10:00 -A <account>


# 8 MPI ranks example
srun python pi_mpi.py

B) Batch job (CPU)

Create run_pi.slurm:

#!/bin/bash
#SBATCH -J mpi-pi
#SBATCH -A <account>
#SBATCH -C cpu
#SBATCH -q debug
#SBATCH -N 1
#SBATCH -t 00:05:00
#SBATCH -o pi-%j.out

module load PrgEnv-gnu python cray-mpich

srun python pi_mpi.py

Submit:

sbatch run_pi.slurm

For longer runs, change -q debug to -q regular and adjust -t appropriately.

---

Notes for GPU queues (optional)

This π example is CPU-oriented. If adapting for GPU nodes:

• Use -C gpu and set proper GPU modules/containers. MPI pattern is unchanged.
• Python+MPI on GPU is typically for orchestrating GPU work per rank (not needed for this simple CPU Monte Carlo demo).

---

License: Educational material for an Intro to HPC Bootcamp.