Processor architectures have grown in leaps and bounds over the last couple of decades, but the software systems used for computational purposes have not kept up with these changes. This paper proposes a new approach to modify legacy source code so that it can be compiled with parallel software compilers that exploit features of the multiprocessor architecture. The review includes a study of some existing software tools like EasyPar, MetaFork, and S2P (used for converting C/C++ sequential code to parallel code), JPar (used to parallelize and optimize Java code), and Casper (for parallelizing sequential Java code to Spark).

Parallelism is highly dependent on the identification of data, control, and resource dependence. If left unresolved, this dependency can lead to indeterminate results. The authors suggest that by analyzing dependency and by generating a dependency graph, an improvised software tool could be developed for enabling parallelism in serial code. The new approach would generate parallel code that suits any of the OpenMP, message passing interface (MPI), and CUDA compilers. The authors find this to be a worthwhile innovation because most other code converters allow transformation that suits only one particular parallel compiler. The paper further suggests that by enabling a web service for this approach, whereby the sequential code and its parallel counterpart can be archived, conversion tasks can also be simplified in the future.

The paper should have mentioned Bernstein’s 1966 classical model on resource dependence, and benchmarking tests involving Amdahl’s law would have provided more credence to the claims made when comparing other similar tools. The work might be of interest to researchers and developers working on parallel software models.