Flare: Native Compilation for Heterogeneous Workloads in Apache Spark
Grégory M. Essertel, Ruby Y. Tahboub, James M. Decker, Kevin J. Brown, Kunle Olukotun, Tiark Rompf
The need for modern data analytics to combine relational, procedural, and map-reduce-style functional processing is widely recognized. State-of-the-art systems like Spark have added SQL front-ends and relational query optimization, which promise an increase in expressiveness and performance. But how good are these extensions at extracting high performance from modern hardware platforms?
While Spark has made impressive progress, we show that for relational workloads, there is still a significant gap compared with best-of-breed query engines. And when stepping outside of the relational world, query optimization techniques are ineffective if large parts of a computation have to be treated as user-defined functions (UDFs). We present Flare: a new back-end for Spark that brings performance closer to the best SQL engines, without giving up the added expressiveness of Spark. We demonstrate order of magnitude speedups both for relational workloads such as TPC-H, as well as for a range of machine learning kernels that combine relational and iterative functional processing.
Flare achieves these results through (1) compilation to native code, (2) replacing parts of the Spark runtime system, and (3) extending the scope of optimization and code generation to large classes of UDFs.