Processing Java UDFs in a C++ environment

Abstract

Many popular big data analytics systems today make liberal use of user-defined functions (UDFs) in their programming interface and are written in languages based on the Java Virtual Machine (JVM). This combination creates a barrier when we want to integrate processing engines written in a language that compiles down to machine code with a JVM-based big data analytics ecosystem. In this paper, we investigate efficient ways of executing UDFs written in Java inside a data processing engine written in C++. While it is possible to call Java code from machine code via the Java Native Interface (JNI), a naive implementation that applies the UDF one row at a time incurs a significant overhead, up to an order of magnitude. Instead, we can significantly reduce the costs of JNI calls and data copies between Java and machine code, if we execute UDFs on batches of rows, and reuse input/output buffers when possible. Our evaluation of these techniques using different scalar UDFs, in a prototype system that combines Spark and a columnar data processing engine written in C++, shows that such a combination does not slow down the execution of SparkSQL queries containing such UDFs. In fact, we find that the execution of Java UDFs inside an embedded JVM in our C++ engine is 1.12X to 1.53X faster than executing in Spark alone. Our analysis also shows that compiling Java UDFs directly into machine code is not always beneficial over strided execution in the JVM.