A Formal Approach to Detect Functionally Irrelevant Barriers in MPI Programs

  • Authors:

  • Subodh Sharma;Sarvani Vakkalanka;Ganesh Gopalakrishnan;Robert M. Kirby;Rajeev Thakur;William Gropp

  • Affiliations:

  • School of Computing, Univ. of Utah, Salt Lake City, USA UT 84112;School of Computing, Univ. of Utah, Salt Lake City, USA UT 84112;School of Computing, Univ. of Utah, Salt Lake City, USA UT 84112;School of Computing, Univ. of Utah, Salt Lake City, USA UT 84112;Math. and Comp. Sci. Div., Argonne Nat. Lab., Argonne, USA IL 60439;Dept. of Computer Sci., Univ. of Illinois, Urbana, Illinois, USA 61801

  • Venue:
  • Proceedings of the 15th European PVM/MPI Users' Group Meeting on Recent Advances in Parallel Virtual Machine and Message Passing Interface
  • Year:
  • 2008

Quantified Score

Hi-index 0.00

Visualization

Abstract

We examine the unsolved problem of automatically and efficiently detecting functionally irrelevant barriers in MPI programs. A functionally irrelevant barrier is a set of MPI_Barriercalls, one per MPI process, such that their removal does not alter the overall MPI communication structure of the program. Static analysis methods are incapable of solving this problem, as MPI programs can compute many quantities at runtime, including send targets, receive sources, tags, and communicators, and also can have data-dependent control flows. We offer an algorithm called Fib to solve this problem based on dynamic (runtime) analysis. Fib applies to MPI programs that employ 24 widely used two-sided MPI operations. We show that it is sufficient to detect barrier calls whose removal causes a wildcard receive statementplaced before or after a barrier to now begin matching a send statement with which it did not match before. Fib determines whether a barrier becomes relevant in anyinterleaving of the MPI processes of a given MPI program. Since the number of interleavings can grow exponentially with the number of processes, Fib employs a sound method to drastically reduce this number, by computing only the relevant interleavings. We show that many MPI programs do not have data dependent control flows, thus making the results of Fib applicable to all the input data the program can accept.