Parallel parsing made practical

Abstract

The property of local parsability allows to parse inputs through inspecting only a bounded-length string around the current token. This in turn enables the construction of a scalable, data-parallel parsing algorithm, which is presented in this work. Such an algorithm is easily amenable to be automatically generated via a parser generator tool, which was realized, and is also presented in the following. Furthermore, to complete the framework of a parallel input analysis, a parallel scanner can also combined with the parser. To prove the practicality of a parallel lexing and parsing approach, we report the results of the adaptation of JSON and Lua to a form fit for parallel parsing (i.e. an operator-precedence grammar) through simple grammar changes and scanning transformations. The approach is validated with performance figures from both high performance and embedded multicore platforms, obtained analyzing real-world inputs as a test-bench. The results show that our approach matches or dominates the performances of production-grade LR parsers in sequential execution, and achieves significant speedups and good scaling on multi-core machines. The work is concluded by a broad and critical survey of the past work on parallel parsing and future directions on the integration with semantic analysis and incremental parsing.