PMR vs STD Containers Benchmark

Performance comparison between std::pmr containers and standard STL containers.

Requirements

• C++17 compiler
• CMake 3.15+
• Python 3.8+ with pandas and matplotlib

pip install pandas matplotlib numpy

Build

Windows:

mkdir build && cd build
cmake ..
cmake --build . --config Release

Linux:

mkdir build && cd build
cmake ..
cmake --build . --config Release

Run

Windows (automated):

.\scripts\build_and_run.ps1

Windows (manual):

cd build\Release
.\benchmark_vector.exe --csv > ..\..\results\raw\vector_results.csv
.\benchmark_string.exe --csv > ..\..\results\raw\string_results.csv
.\benchmark_map.exe --csv > ..\..\results\raw\map_results.csv
.\benchmark_mixed.exe --csv > ..\..\results\raw\mixed_results.csv

Linux:

cd build
./benchmark_vector --csv > ../results/raw/vector_results.csv
./benchmark_string --csv > ../results/raw/string_results.csv
./benchmark_map --csv > ../results/raw/map_results.csv
./benchmark_mixed --csv > ../results/raw/mixed_results.csv

Generate Reports

python scripts/generate_report.py
python scripts/plot_results.py

Key Findings

From running these benchmarks:

Performance Trade-offs

• PMR is 3-4x slower for basic operations (vector push_back, string concat)
• PMR offers 2-3x better determinism (lower variance/CV in timing)
• Pool allocator shows best consistency (CV as low as 5.6%)

When to Use PMR

• Hard real-time systems where consistent timing beats average speed
• Embedded systems that avoid heap allocation
• Safety-critical software that needs deterministic behavior
• Short-lived operations where you know the memory limits upfront

When to Use std::allocator

• When you need maximum speed
• Workloads that already have consistent timing
• When timing isn't critical and performance matters most

Blog Post

For the full analysis and all benchmark results, check out the blog post:
PMR vs std::allocator: When 3x Slower is Better

It covers:

• Deep dive into performance numbers
• Charts showing the trade-offs
• Real code examples
• Memory resource comparisons
• Guidelines for choosing the right allocator

What's Tested

• Vector: push/pop, random access, sorting
• String: concatenation, copying, SSO behavior
• Map: insert/lookup/delete with int and string keys
• Mixed: sensor data collection, message queues

Memory Resources

• std::allocator - standard allocator
• std::pmr::monotonic_buffer_resource - no deallocation
• std::pmr::unsynchronized_pool_resource - pooled allocation