npynuma

NUMA node allocation for NumPy arrays on Linux

Install System Requirements

# Ubuntu/Debian
sudo apt install libnuma-dev

# Fedora/CentOS
sudo dnf install numactl-devel

# Python package
git clone https://github.com/xsgeng/npynuma
cd npynuma
pip install .

pip install npynuma # not yet

Quick Start

import numpy as np
from npynuma import NumaPolicy

with NumaPolicy(node=1):  # Allocate on NUMA node 1
    large_array = np.empty(10_000_000)  # 80MB array on node 1

# Works with existing numpy functions
with NumaPolicy(node=0):
    random_data = np.random.normal(size=1_000_000)

Notes

• Requires Linux NUMA system (numactl installed)
• Node numbers must exist on your system

Thread Safety

⚠️ Not thread-safe - The NUMA policy context is global to the process. Avoid using NumaPolicy contexts concurrently in:

• Multithreaded applications
• Async frameworks
• Any parallel code sharing NumPy allocations