SIMD for Java: how Elasticsearch already benefits from the Panama Vector API

Processing large amounts of data efficiently is a common challenge in modern systems. Many workloads spend most of their time performing simple mathematical operations on large arrays of numbers. Modern CPUs can speed this up using SIMD instructions, which allow a single instruction to process multiple values at once. Traditionally, accessing these capabilities from Java has been difficult and often required platform-specific native code.

The Panama Vector API changes this by allowing Java developers to express these operations directly in Java while the JVM turns them into efficient CPU instructions.

Elasticsearch provides an interesting real-world example. Modern features such as vector search rely on large numbers of vector calculations, and analytical queries written in ES|QL often process large batches of numeric data. In both cases the system repeatedly performs simple operations over large arrays of values. Elasticsearch and its underlying search engine Lucene use the Vector API to speed up these operations by taking advantage of SIMD instructions available on modern CPUs.

Using an incubating API in production software is unusual and normally avoided. In this case, however, the potential performance gains are significant enough to make it worthwhile. This talk explores how Elasticsearch integrates the Vector API, how it safely enables and disables SIMD acceleration at runtime, and what this tells us about the future of high-performance computing in Java.