Parallel XML Transformations on Multi-Core Processors

Abstract

With the advent of multicores and XML, we face the mounting need of processing XML documents efficiently on multicores. Because XSLT (extensible stylesheet language for transformation) is the principal programming language to transform one XML document into another, we design and implement a novel and powerful XSLT engine based on multicores. Since XSLT is declarative and functional, implicit parallelism abounds in XSLT processing. This paper addresses the challenges to map such parallelism to various execution models and to exploit these mappings with optimal granularities. We develop four parallel execution models for XML transformations: fork, fork-join, data-parallel, and future. These models are our building blocks for data and computation decompositions unto multicores. Not only are we the first to formulate them on top of the building blocks, we also demonstrate how to efficiently exploit implicit, dynamic and nested parallelism with both compile-time estimation and profile-based adaptation on granularity. Finally, we present the first-ever evaluation of a parallel XSLT engine on multicore processors. The performance analysis shows that our parallel XSLT engine is scalable for most XSLT programs in real-world usage.