Locating Parallelization Potential in Object-Oriented Data Structures

Abstract

The free lunch of ever increasing single-processor performance is over. Software engineers have to parallelize software to gain performance improvements. But not every software engineer is a parallel expert and with millions of lines of code that have not been developed with multicore in mind, we have to find ways to assist in identifying parallelization potential. This paper makes three contributions: 1) An empirical study of more than 900,000 lines of code reveals five use cases in the runtime profile of object-oriented data structures that carry parallelization potential. 2) The study also points out frequently used data structures in realistic software in which these use cases can be found. 3) We developed DSspy, an automatic dynamic profiler that locates these use cases and makes recommendations on how to parallelize them. Our evaluation shows that DSspy reduces the search space for parallelization by up to 77% and engineers only need to consider 23% of all data structure instances for parallelization.