Petal tool for analyzing and transforming legacy MPI applications

Abstract

Legacy MPI applications are an important and economically valuable category of parallel software that rely on the MPI-1, MPI-2 (and, more recently, MPI-3) standards to achieve performance and portability. Many of these applications have been developed or ported to MPI over the past two decades, with the implicit (dual) goal of achieving acceptably high performance and scalability, and a high level of portability between diverse parallel architectures. However they were often created implicitly using MPI in ways that exploited how a particular underlying MPI behaved at the time (such as those with polling progress and poor implementation of some operations). Thus, they did not necessarily take advantage of the full potential for describing latent concurrency or for loosening the coupling of the application thread from the message scheduling and transfer. This paper presents a first transformation tool, Petal, that identifies calls to legacy MPI primitives. Petal is implemented on top of the ROSE source-to-source infrastructure and automates the analysis and transformation of existing codes to utilize non-blocking MPI and persistent MPI primitives. We use control flow and pointer alias analysis to overlap communication and computation. The transformed code is capable of supporting better application bypass, yielding better overlapping of communication, computation, and I/O. We present the design of the tool and its evaluation on available benchmarks.