Memory Hierarchy Experiment Framework

A Linux benchmarking framework for studying memory bottlenecks, cache locality, TLB-related behavior, and matrix blocking effects using C microbenchmarks and perf-based automation.

Overview

This project was built to analyze how memory-access patterns and working-set sizes affect performance on Linux systems. It combines:

• C microbenchmarks
• Bash automation scripts
• Python parsing and plotting tools
• perf-based performance measurement

Core Experiments

1. Memory access patterns

• sequential access
• strided access
• random access

2. Matrix multiplication variants

• naive loop order
• reordered loop order
• blocked / tiled execution

3. Virtual memory behavior

• page-level sequential access
• page-level random access

Repository Structure

src/        benchmark implementations
scripts/    build, collection, parsing, plotting automation
docs/       workflow and reproducibility notes
report/     final write-up and figures
tests/      smoke test
archive/    non-essential supporting materials kept out of the main path

Build

make

Run

Collect perf results:

./scripts/collect_perf.sh

Parse one run manifest:

python3 ./scripts/parse_perf.py \
  --manifest results/raw/<run-id>/manifest.csv \
  --output results/processed/perf_results.csv

Generate plots:

python3 ./scripts/plot_results.py \
  --input results/processed/perf_results.csv \
  --plots-dir plots

If hardware PMU events are unavailable, use:

./scripts/collect_bench_only.sh

Key Focus

• cache locality and working-set transitions
• memory bottlenecks under different access patterns
• TLB-related behavior at page granularity
• matrix blocking effects on cache reuse

Result Summary

This framework enables reproducible comparison of memory-access patterns, matrix traversal orders, and page-level access behavior using Linux perf counters such as cycles, instructions, cache misses, page faults, and dTLB-load-misses.

Notes

• The main artifacts intended for review are under src/, scripts/, docs/, and report/.
• Supporting materials that are not essential to the core benchmarking workflow were moved under archive/ to keep the repository focused.