EmbeddingGemma 2: an open, lightweight multimodal embedding model
deepmind.google·6d ago
AI Engineering
Hacker News·Aug 14·5 min read
High-Performance Array-Backed LRU Hash Table
TL;DR
A high-performance concurrent LRU hash table was developed for systems programming, achieving over 14x throughput and sub-microsecond latencies by using custom spinlocks and a sharded architecture.
✦ Why It Matters
Consider implementing this LRU hash table in your caching layer to enhance performance under load.
Key Takeaways
- Implement sharded architecture to eliminate global lock contention.
- Use pre-allocated flat arrays for better cache performance.
- Adopt lazy LRU promotion to improve read throughput.
- Benchmark performance on target hardware to optimize configurations.
How It Works
The LRU hash table is divided into independent shards, each with its own spinlock and memory allocation. This design prevents global lock contention and allows for parallel operations.
The use of 32-bit indices instead of 64-bit pointers reduces memory overhead and enhances cache locality. Additionally, the lazy LRU promotion mechanism probabilistically skips promotions on hot reads, allowing for faster access without locking.