A large-scale study of MPI usage in open-source HPC applications

Abstract

Understanding the state-of-the-practice in MPI usage is paramount for many aspects of supercomputing, including optimizing the communication of HPC applications and informing standardization bodies and HPC systems procurements regarding the most important MPI features. Unfortunately, no previous study has characterized the use of MPI on applications at a signicant scale; previous surveys focus either on small data samples or on MPI jobs of specic HPC centers. This paper presents the rst comprehensive study of MPI usage in applications. We survey more than one hundred distinct MPI programs covering a signicantly large space of the population of MPI applications. We focus on understanding the characteristics of MPI usage with respect to the most used features, code complexity, and programming models and languages. Our study corroborates certain ndings previously reported on smaller data samples and presents a number of interesting, previously unreported insights. CCS CONCEPTS• General and reference → Surveys and overviews; • Computing methodologies → Parallel programming languages.

Identifying Degree and Sources of Non-Determinism in MPI Applications Via Graph Kernels

“…These constructs are typical of non-deterministic application's communication patterns, as presented in [1], [2], [7]. Furthermore, characterization of trends in how HPC applications use MPI features presented in [8] proves that the majority of MPI applications with nondeterministic patterns use point-to-point communication constructs, guaranteeing the broader impact of our framework across scientific domains. Thus, the value proposition of our work is that, supported by our framework, scientists can ultimately bootstrap their process for diagnosing nondeterministic bugs or addressing reproducibility failures reducing the labor-intensive manual inspection of complex code.…”

Towards compiler-aided correctness checking of adjoint MPI applications

“…A recent study of about 100 MPI applications [24] has shown, that most codes require MPI-2 or lower and point-topoint communication and collectives are the majority of used MPI features. However, derived MPI datatypes are the fourth most used feature in this study.…”

The MPI Bugs Initiative: a Framework for MPI Verification Tools Evaluation

“…It is worth mentioning that some features are irrelevant to some errors (e.g., RMA operations cannot produce a Request Lifecycle error). It may be observed that all MPI features are represented in the benchmark with a predominance of basic collectives and point-to-point operations which correspond to the most used features in HPC applications [30].…”