Nice Bench

This Python script benchmarks how different nice levels affect CPU-bound tasks, measuring throughput and latency. It runs processes with varying nice levels to observe their impact on execution time.

Requirements

• Python 3.x
• A Unix-like OS with nice support (e.g., Linux, macOS)

Usage

Command-line arguments:

1. -t, --task-size: CPU work per task, in calibrated seconds of single-core work (default: 9).
2. -n, --nice-duplicates: Number of processes per nice level (default: 2).
3. -o, --output: Telemetry JSON path (default: nice-bench-<timestamp>.json).
4. --default-timer-slack: Keep the kernel's default timer slack (usually 50us). By default tasks set it to 1ns so wake-up latency reflects scheduling rather than slack.

Output

All tasks wait at a barrier, then wake from a timed sleep at one common start time, so the scheduler alone decides who runs first. Each task sleeps 1ms per ~10ms of CPU work to sample wake-up latency. At the end a table shows, per task: finish time and rank, CPU time and runqueue wait (from /proc/self/schedstat), involuntary context switches, latency p50/p90/p99/max, and a finish-time timeline. !! marks latency outliers (more than 5 median absolute deviations above the median).

The summary reports how well finish order follows nice order (Kendall tau, 1.0 = perfect), the number of inversions and the worst one, inversions broken down by nice distance (to tell equal-priority coin flips from real misordering), and the latency extremes across all samples.

The JSON file holds run metadata and every task's telemetry, including each wake-up sample's timestamp, latency and CPU, for offline analysis.

Example Usage

Default run:

sudo python3 nice_bench.py

Custom run without duplicate processes and reduced task size:

sudo python3 nice_bench.py -t 7 -n 1

Note: Invocation with sudo is required to set nice level for restricted nice levels.