Workshops Archive

An LLVM Pass for Automatic Skeletonization of MPI Applications


Workshop: LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC

Authors: Quentin Buot (Inria); Emmanuel Jeannot (National Institute for Research in Computer Science and Automation (Inria), France; University of Bordeaux); Pierre Lemarinier (Bull); Guillaume Mercier (Bordeaux INP); and Richard Sartori (INRIA)

Abstract: We present an LLVM pass that automatically skeletonizes MPI applications: it statically identifies computations that do not affect the arguments of MPI calls and removes them, producing a much faster program that preserves the original's communication semantics. The pass operates on the LLVM IR, propagating a communication variable status backward from MPI call sites through the program's data and control dependencies, and handles the non-trivial engineering challenges (recursion, loops, dynamic libraries, memory allocation) that arise when applying this analysis to real, non-trivial HPC codes such as Lulesh and miniFE. We validate that the generated skeletons reproduce the communication pattern of the original application, and illustrate one concrete use of this technique: accelerating MPI runtime tuning by running the parameter search on the skeleton rather than on the full application.


Back to LLVM-HPC2026: The Twelfth Workshop on the LLVM Compiler Infrastructure in HPC Archive Listing

Back to Full Workshop Archive Listing