Vector search performance of Charmed OpenSearch: the benefits of Intel® AVX-512
Accelerating OpenSearch operator in Intel® Advanced Vector Extensions 512
About the guide
The use of vector databases is rapidly growing due to the increasing volume of data and the expansion of vector dimensions, such as 128, 256, 768, 1536, and 2048. However, a significant challenge in such systems is the time required to build a vector index and conduct vector searches. This process involves mathematical operations such as Euclidean distance, cosine similarity, and dot product similarity to compute the similarity between vectors. As datasets and their dimensionality increase, these calculations become more computationally intensive, posing unique challenges for the designers of vector databases and necessitating the exploration of additional ways to speed up vector operations.
This whitepaper explores how hardware acceleration impacts the performance of vector operations within Charmed OpenSearch running on Intel® Xeon® processors with AVX-512 vector acceleration support. Enabling and disabling hardware acceleration for the benchmarks is performed through the OpenSearch Vector API.
We will provide insights and recommendations for optimising vector search operations in large-scale applications by improving latency and throughput using hardware accelerators for vector capabilities.
In this context, throughput refers to OpenSearch’s ability to handle more documents per second. However, this can be a resource-intensive task in vector databases, as incoming data needs to be processed to determine its nearest neighbours. Similarly, search latency is affected as the search query must be processed to identify the closest data points in the database.
We will present Charmed OpenSearch’s throughput and latency advancements to demonstrate that technologies like AVX-512 can enhance the baseline.