Moreton-Fernandez et al. present an approach to automatically determine communication patterns at runtime for programs that run on distributed-memory parallel systems. Optimizing and tuning code for such systems when using traditional, low-level programming models such as message-passing interface (MPI) is still a complex problem even for experienced developers.

The suggested approach enables efficient adaption of a program’s communication and synchronization, as well as optimizations at runtime. The approach is implemented in the Trasgo compiler framework that compiles a high-level parallel language for hierarchical systems into an Extensible Markup Language (XML) internal representation (IR).

The IR is then transformed to C code, which uses the Hitmap runtime library for an efficient data distribution and for aggregating communication; the C code is finally compiled using the Pluto polyhedral compiler.

The presented approach aims to liberate application developers from cumbersome, low-level design tasks by providing a high-level and easy-to-use domain-specific language (DSL) for parallel programming. Since performance-critical parameters, such as tile size, may significantly influence communication patterns, the approach shifts the choice of parameter values from compile-time to runtime and thus enables adaption flexibility and portability even for heterogeneous parallel systems with nodes that have different capabilities. Therefore, this approach is especially interesting for application programmers that aim to optimize communication patterns in their programs for heterogeneous high-performance systems.

The authors demonstrate promising results in a detailed experimental evaluation, which also includes measuring the overhead for the runtime adaption of parameters. Unfortunately, experimental results are presented in large tables instead of plots, which makes the comparison of achieved results quite laborious.