Contributed Articles Formal Analysis of Mpi-based Parallel Programs

MosT ParaLLeL CoMPUTiNG applications in highperformance computing use the Message Passing Interface (MPI) API. Given the fundamental importance of parallel computing to science and engineering research, application correctness is paramount. MPI was originally developed around 1993 by the MPI Forum, a group of vendors, parallel programming researchers, and computational scientists. however, the document defining the standard is not issued by an official standards organization but has become a de facto standard through near-universal adoption. MPI development continues, with MPI-2.2 released in 2009 and MPI 3.0 expected in 2012. The standard is published on the Web and as a book, along with several other books based on it; see, for example, Gropp et al. and Pacheco.26 Implementations are available in open source from MPICH2 and Open MPI25 from software vendors and from every vendor of HPC systems. MPI is widely cited; Google Scholar recently returned 39,600 hits for the term “+MPI +Message Passing Interface.” MPI is designed to support highly scalable computing applications using more than 100,000 cores on, say, the IBM Blue Gene/P (see Figure 1) and Cray XT5. Many MPI programs represent dozens, if not hundreds, of personyears of development, including calibration for accuracy and performance tuning. Scientists and engineers worldwide use MPI in thousands of applications, including in investigations of alternate-energy sources and in weather simulation. For HPC computing, MPI is by far the dominant programming model; most (at some centers, all) applications running on supercomputers use MPI. Many application developers for exascale systems regard support for MPI as a requirement. Still, the MPI debugging methods available to these developers are typically wasteful and ultimately unreliable. Existing MPI testing tools seldom provide coverage guarantees, examinformal analysis of mPI-based Parallel Programs doI:10.1145/2043174.2043194