cpp-performance-lab

A focused C++ microbenchmark repository for cache, memory, ILP, synchronization, queue, and allocator behavior.

Purpose

• Build intuition for cache hierarchy and memory access patterns
• Compare synchronization and data-structure tradeoffs with reproducible microbenchmarks
• Produce evidence-based performance notes from stable runs

1) Prerequisites

• CMake >= 3.20
• A C++20 compiler (clang++ or g++)
• Git + internet access (for fetching google/benchmark)

2) Build

cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j

3) Run all benchmarks (recommended)

scripts/run_all.sh

4) Run a specific benchmark

Standard pattern:

./build/benchmark/<binary_name> --benchmark_min_time=0.3s

Examples:

./build/benchmark/bm_stride_access --benchmark_min_time=0.3s
./build/benchmark/bm_cache_levels --benchmark_min_time=0.3s

Queue tuned run:

./build/benchmark/bm_queue \
  --benchmark_filter='BM_Queue(MutexTransfer/batch:64/backoff:0|SpscRingTransfer/batch:8/backoff:0)$' \
  --benchmark_min_time=1s \
  --benchmark_repetitions=10 \
  --benchmark_report_aggregates_only=true

5) Benchmarks

• bm_stride_access: locality loss from larger access stride
• bm_pointer_chasing: sequential access vs irregular pointer traversal
• bm_false_sharing: adjacent counters vs cache-line-padded counters
• bm_aos_vs_soa: layout sensitivity for dense vs sparse field usage
• bm_mutex_vs_atomic: contention scaling for shared counter updates
• bm_cache_levels: throughput drop as working set crosses cache levels
• bm_ilp: dependent vs independent instruction streams
• bm_cache_associativity: friendly stride vs conflict-prone stride
• bm_queue: mutex queue vs tuned SPSC ring transfer
• bm_memory_pool: new/delete vs locked pool vs thread-local pool

6) Outputs

• results-summary.md: current run summary and conclusions

Generate figures from benchmark runs:

python3 scripts/generate_plots.py

7) Figures

Cache Levels

Cache Associativity

Queue and Memory Pool