DEV Community

CharmPic
CharmPic

Posted on

Hakozuna v3.2 Released: Bringing Optimized Memory Allocation to M1 Mac

I've added the main text to chatgpt5.4 ↓

I am pleased to announce the release of Hakozuna v3.2.
While my previous update focused on Windows, this release marks a significant milestone: Full support for M1 Mac.

GitHub Release:https://github.com/hakorune/hakozuna

Zenodo Record: 19120414

DOI: 10.5281/zenodo.19120414

What is Hakozuna?
Hakozuna is a memory allocator designed for small objects, built upon the Box Theory framework. It is currently split into two specialized lineages:

hz3: Optimized for local-heavy / low-RSS workloads.

hz4: Optimized for remote-heavy / high-thread environments.

What’s New in the M1 Mac Update
The primary goal of this update was to establish a seamless workflow on M1 Mac—encompassing development, observation, and running benchmarks for academic papers.

Key Improvements:

Refined Mac Entrypoints: All Mac-specific logic is now consolidated in the mac/ directory.

Pipeline Separation: Decoupled the Build Lane and Observe Lane for better modularity.

Streamlined Paper-Suite: You can now run the full suite of benchmarks required for research papers with a single setup.

Comparative Benchmarking: Integrated mimalloc and tcmalloc into the suite to allow direct performance comparisons against hz3 and hz4.

Performance Insights: Where it Shines
Testing the paper-suite on Mac revealed clear strengths for each allocator:

hz3 showed dominant performance in the Larson benchmark.

hz4 took the lead in MT remote (Multi-threaded remote free) scenarios.

In Redis-like workloads, the winner shifted depending on the specific workload characteristics.

Note on mimalloc-bench: In our subset tests, certain malloc-large treatments were flagged as "no-go."

Segment Registry: For high-remote conditions, we found that slots=32768 yielded better results.

The Takeaway:
The M1 Mac results reinforce our core philosophy: rather than trying to create a "one-size-fits-all" allocator, partitioning "boxes" based on specific conditions leads to superior efficiency.

Conclusion
With v3.2, "Mac support" is more than just a port—it is a functional environment ready for rigorous academic benchmarking.

Summary of Gains:

Improved DX (Developer Experience) on M1 Mac.

Automated and reliable comparative benchmarking via paper-suite.

Clearer functional boundaries between the hz3 and hz4 lineages.

Next, I plan to utilize this Mac environment to refine the supplementary data and further validate my research for the upcoming paper.

Top comments (0)