A Performance and Scalability Analysis of the MPI Based Tools Utilized in a Large Ice Sheet Model Executing in a Multicore Environment

Abstract

This paper analyzes the performance and scalability characteristics of both the computational and I/O components of the Parallel Ice Sheet Model (PISM) executing in a multicore supercomputing environment. It examines the impact of multicore technologies on two state-of-the-art parallel I/O systems, both of which are based on the same underlying implementation of the MPI-IO standard, but which exhibit very different performance and scalability characteristics. It also examines these same characteristics for the MPI-based computational engine of the simulation model. One important benefit of studying these three software systems both independently and together is that it exposes a fundamental tradeoff in the ability to provide scalable I/O and scalable computational performance in a multicore environment. This paper also provides what, at least at first glance, appears to be very counter-intuitive performance results. We examine the underlying reasons for such results, and discuss the important insights gained through this examination.