SC14 New Orleans, LA

The International Conference for High Performance Computing, Networking, Storage and Analysis

Static Analysis of MPI Programs Targeting Parallel Properties.

Student: Sriram Aananthakrishnan (University of Utah)
Supervisor: Ganesh Gopalakrishnan (University of Utah)

Abstract: MPI is one of the dominant programming model for writing HPC applications. Unfortunately, debugging MPI programs is hard. It is well known that static dataflow analysis discovers provably true properties that are useful in optimization and proving correctness. Existing techniques fail to interpret MPI operations for their message passing semantics thereby lacking the ability to recognize parallel properties of MPI programs. Parallel properties are properties about communication topology or properties that depend on communication; e.g., reaching constants over MPI operations. In this work we propose (i) new abstractions and techniques to statically discover the communication topology (ii) provide a framework for writing dataflow analyses to recognize parallel properties of MPI programs.

Poster: pdf
Two-page extended abstract: pdf

Poster Index