Parallel Parsing: The Earley and Packrat Algorithms

Parsing plays a critical role in our modern computer infrastructure: scripting languages such as Python and JavaScript, layout languages such as HTML, CSS, and Postscript/PDF, and data exchange languages such as XML and JSON are all interpreted, and so require parsing. Moreover, by some estimates, the time spent parsing while producing a rendered page from HTML, CSS, and JavaScript is as much as 40%. Motivated by the importance of parsing, and also by the wide-spread adoption of chip multiprocessors, we are interested in investigating the potential for parallel parsing algorithms. We begin with one classical and one more modern parsing algorithm, the Earley and the Packrat algorithm, respectively, and consider methods for parallelizing both. The general strategy is the same in both cases: break the string to be parsed into contiguous blocks, process each block in parallel, and finally re-join the blocks. We find that using this general framework, we are able to obtain a meaningful speedup relative to serial implementations of each algorithm. In particular, we obtain a speedup of 5.4 using the parallel Earley implementation on 8 processors, and a speedup of 2.5 using the parallel Packrat algorithm on 5 processors. Another algorithmic property that we consider is work efficiency, a measure of the amount of extra work done by the parallel algorithm relative to the serial algorithm. Poor work efficiency indicates that the parallel algorithm is doing considerably more work than is required, and thereby using more energy. With the relative importance of energy increasing, even a parallel algorithm that achieves significant speedup may not be worthwhile if it exhibits poor work efficiency. Though it is difficult to formulate a precise definition of work efficiency in terms of program operations, we qualitatively demonstrate that our algorithms, with some caveats, exhibit reasonable work efficiency.