Milvus KScaNN Algorithm Optimization
Application scenario: High performance with high precision is required.
Technical principle: Scalable Nearest Neighbors (ScaNN), an open source algorithm library released by Google for efficient vector similarity search, leverages the principles of IVF-PQ to achieve outstanding retrieval performance. This is accomplished through x86-optimized 4-bit SIMD processing and enhanced anisotropic quantization loss functions. Kunpeng Scalable Nearest Neighbors (KScaNN) represents the Arm-adapted version of this technology, featuring architecture-specific optimizations for Kunpeng processors. By re-engineering index structures, algorithmic workflows, and computational processes around inverted indexes, KScaNN maximizes hardware potential through Kunpeng's vectorization instructions and advanced quantization techniques, offering search capabilities on par with the open source ScaNN implementation.
Performance metric: Compared with the Milvus-ScaNN algorithm, the Milvus-KScaNN algorithm improves QPS performance by 30% on the ANN-Benchmarks GIST dataset with a recall value greater than 0.99 and the configuration of 16 vCPUs and 64 GB memory.
