An empirical examination of the prevalence of inhibitors to the parallelizability of open source software systems

Abstract

An empirical study is presented that examines the potential to parallelize generalpurpose software systems. The study is conducted on 13 open source systems comprising over 14 MLOC. Each for-loop is statically analyzed to determine if it can be parallelized or not. A for-loop that can be parallelized is termed a free-loop. Free-loops can be easily parallelized using tools such as OpenMP. For the loops that cannot be parallelized, the various inhibitors to parallelization are determined and tabulated. The data shows that the most prevalent inhibitor by far, is functions called within for-loops that have side effects. This single inhibitor poses the greatest challenge in adapting and re-engineering systems to better utilize modern multi-core architectures. This fact is somewhat contradictory to the literature, which is primarily focused on the removal of data dependencies within loops. Results of this paper also show that function calls via function pointers and virtual methods have very little impact on the for-loop parallelization process. Historical data over a 10-year period of inhibitor counts for the set of systems studied is also presented. It shows that there is little change in the potential for parallelization of loops over time.

Middleware and multicore architecture: Challenges and potential enhancements from software engineering perspective

“…At the onset, this may appear to be a problematic, however conservative, limitation. However, this assumption is supported by empirical analysis we undertook in a previous study [7].…”

“…A for-loop is considered a free-loop if it does not contain any parallelization inhibitors that are not already solvable with OpenMP. We used a tool, ParaStat, developed by one of the main authors and used in [7], to analyze loops and determine if they contain any inhibitors as defined in this section. First, we collected all files with C/C++ source-code extensions (i.e., c, cc, cpp, cxx, h, and hpp).…”

“…S. Alnaeli, J. Maletic et al [7] conducted an empirical study that examines the potential to parallelize large-scale generalpurpose software systems. They found that the greatest inhibitor to automated parallelization of for-loops is the presence of function calls with side effects and they empirically proved that this is a common trend.…”

“…All methods are conservative in the case of dependency suspension, or when it is difficult to prove the opposite, so that no unsafe parallel transformation is done. [7] = constRef [ The main purpose of data-dependency analysis is to detect if the same memory position is used in more than one loop iteration. The majority of dependency analysis algorithms are focused on array references.…”

“…The auto-parallelization can also take place via a tool prior to compiling. These tools look for for-loops that do not contain any parallelization inhibitors [6,7].…”