EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Lobsters·24d ago·5 min read
Benchmarking Wild vs Mold
TL;DR
Mold's recent linker benchmarks reveal that Wild is significantly slower than Mold, prompting an investigation into the discrepancies. Key differences in benchmarking configurations and filesystem choices contribute to the performance gap.
✦ Why It Matters
Consider adjusting your linker benchmarking configurations to align with Mold's practices for more accurate comparisons.
Key Takeaways
- Run benchmarks with output file deletion for accurate results.
- Use ext4 filesystem instead of tmpfs for better performance.
- Implement --no-fork option to reduce shutdown costs.
- Explore OS-specific optimizations to enhance linker performance.
How It Works
Mold's performance benefits from using techniques like fallocate to pre-allocate file space and hugepages for efficient memory mapping, which Wild has yet to fully implement.