This post details the implementation and optimization of a static search tree (S+ tree) for high-throughput searching of sorted data. The primary goal is to significantly improve search performance over traditional binary search, particularly for large datasets.

The Eytzinger layout is introduced as a baseline optimization, reordering data in memory to improve cache efficiency. This layout allows for prefetching multiple cache lines ahead, effectively hiding memory latency. Benchmarks show Eytzinger is up to 6x faster than standard binary search for large datasets, achieving around 200ns/query compared to binary search's 1150ns/query at 1GB input size.

The article then delves into further optimizations to push performance limits. These include techniques like auto-vectorization, manual SIMD instructions (AVX2), batching queries, prefetching, pointer arithmetic, and optimizing the tree layout. The aim is to shave off every possible instruction cycle. A key addition is batching, which significantly boosts throughput. The implementation draws inspiration from prior work on Algorithmica and focuses on maximizing query rates for static data, relevant for applications like bioinformatics indexing.

Ultimately, the developed S+ tree demonstrates a 40x speedup compared to binary search. The source code and benchmarks are available on GitHub. The post concludes by discussing future work, including branchy search, interpolation search, and range queries.