Framework for Lifecycle Enrichment of HPC Applications Towards Exascale Heterogeneous Architectures
Karan Sapra · TigerPrints (Clemson University) · 2018
With the advent of accelerators and architectures, researchers are faced with a daunting task to port their existing applications and algorithms to the optimal architecture and programming language. Porting existing applications or a new algorithm is both demanding and time-consuming due to the sheer number of accelerators and architectures plus the number of programming models available per architecture. This problem is further compounded for heterogeneous systems with wide availability of resources and complexity of scientific applications. In this dissertation, we focus on enriching the lifecycle of applications by providing an application to optimal architecture mapping and framework to assist in making the most effective use of resources in a heterogeneous environment. Our Application to Architecture (A2A) framework can be further divided into sub mappings: Qualitative and Quantitative. Our qualitative mapping uses benchmark application analysis to understand the application performance without in depth runtime analysis and is highly suitable for new algorithms and applications. Our quantitative mapping can provide detail numerical performance analysis for porting an application across programming models and architectures. We evaluate our overall framework using various diverse benchmark applications. Lastly, our Heterogeneous Partitioning Framework (HFP) provides existing and new applications the ability to use heterogeneous resources in an efficient manner with minor modifications to the source code in comparison to other frameworks as is shown with two case studies: Climate Earth Science Model (CESM) and GPU-based Gene Network Alignment Tool (G3NA).