CPU Quotas (soma login nodes)

The soma login nodes are shared resources. The total CPU resources are soft-limited in order to support multiple users. This is done using systemd quotas.

Unfortunately, Python libraries and command line utilities like nproc don't know about the soft limits and provide inaccurate user CPU core values, leading to oversubscribe processes when running custom parallel scripts.

Workaround

In this repository you will find some useful scripts that get the user CPU qouta values for soma login nodes:

1. python (get_available_cpus.py)

This script supports multiple usage modes and options:

Features:

• Reads systemd CPU quota from /run/systemd/system/user-$UID.slice.d/50-CPUQuota.conf
• Handles both percentage format (e.g., "200%") and decimal format (e.g., "2.0")
• Provides a conservative mode that reserves CPU for system overhead
• Can be imported as a module in other Python scripts
• Includes a quiet mode ideal for scripting

Usage:

# Make executable
chmod +x get_available_cpus.py

# Basic usage
./get_available_cpus.py

# Conservative mode (reserves some CPU for system)
./get_available_cpus.py --conservative

# Quiet mode (just the number)
./get_available_cpus.py --quiet

In python scripts:

from get_available_cpus import get_parallel_workers

# Use in multiprocessing
import multiprocessing as mp
num_workers = get_parallel_workers()
with mp.Pool(num_workers) as pool:
    # your parallel code here

2. bash (get_available_cpus.sh)

This is a bash version of the script above for command-line usage:

Usage:

# Make executable
chmod +x get_available_cpus.sh

# Basic usage
./get_available_cpus.sh

# Use in other scripts
NUM_CPUS=$(./get_available_cpus.sh --quiet)

3. pip installing cpu_quota

How to install:

python -m pip install cpu_quota

Usage:

from cpu_quota import get_parallel_workers
num_workers = get_parallel_workers()

Features:

• Systemd Integration: Reads actual CPU quota limits rather than just hardware CPU count
• Fallback: Falls back to nproc if no systemd limits are found
• Conservative Mode: Optionally reserves CPU capacity for system overhead
• Error Handling: Gracefully handles missing files or parsing errors
• Flexible Output: Quiet mode for scripting, verbose mode for debugging