A new proxy application for compressible multiphase turbulent flows

Abstract

CMT-nek is a simulation code for performing high fidelity, predictive simulations of particle laden explosively dispersed turbulent flows under conditions of extreme pressure and temperature. CMT has many applications including lithotripsy, dust explosions in coal mines, and explosive detonation. The physical processes underlying these applications are complex and cover a very wide range of spatial and temporal scales.The underlying workflow of this application requires a very rich set of computation on multiple types of data structures and data access patterns. Additionally, the simulation of such codes require high spatial and temporal resolution resulting in large memory footprint and enormous computing power that is only available on extant petascale and exascale supercomputers. The workflow and the underlying patterns are representative of other scientific applications and hence have the potential to serve as an important benchmark for next generation supercomputers.In this paper,11A preliminary version of this work will appear in Proceedings of HiPC 2016, “CMT-bone: A Proxy Application for Compressible Multiphase Turbulent Flows”, by Tania Banerjee, Jason Hackl, Mrugesh Shringarpure, Tanzima Islam, S. Balachandar, Thomas Jackson and Sanjay Ranka. we describe CMT-bone that serves as a proxy-app for CMT-nek. Proxy applications provide a simplified framework for computational scientists and system designers to investigate new node architectures, programming approaches, algorithmic design choices and optimization techniques. CMT-bone retains the workflow of CMT-nek and encapsulates key data structures, computational and communication patterns of CMT-nek. It reduces the number of variables used and corresponding compute and communication operations performed at each step of the work.Validation of CMT-bone with the CMT-nek using the veritas tool developed at LLNL shows that CMT-nek is a good proxy. Such a validation is important as it shows that important computational characteristics of CMT-nek are embedded in CMT-bone and any experimental performance analysis of the latter will be representative of the former.