CAREER: Unifying Heterogeneity of Extreme-Scale Cyberinfrastructures for Higher Productivity and Performance Portability

点击翻译按钮获取中文摘要

The goal of this project is to improve the performance portability and productivity (ease-of-use) of application development for extreme-scale heterogeneous reconfigurable architectures. This goal is in the scope of NSF's 10 Big Ideas, i.e., "Harnessing the Data Revolution", and a recent national report, i.e., "Extreme Heterogeneity 2018". The project will enable scientists in domains such as quantum simulation, experimental High-Energy Physics (HEP), and Fossil Energy (FE) to focus on the inherent characteristics, e.g., parallelism, of their applications rather than be burdened by the technological details of the underlying hardware. The project also significantly increases the longevity of domain scientists' legacy code. This project integrates the PI's research and teaching activities. The outcome from the project will produce undergraduate and graduate students who can contribute to solving problems in large-scale cyberinfrastructure. The project aims to include women and minority groups, through The Office of Diversity, Equity and Inclusion at the University of Kansas (KU), to receive training and engage in research collaboration with our group. This project will enable higher productivity in post-exascale (extreme-scale) heterogeneous architectures, advancing the PI's long-term research goals. These project goals align with: (1) the mission of the Office of Advanced Cyberinfrastructure (OAC) of NSF and the mission of NSF at large in promoting the progress of science and advancing the national prosperity and welfare, and (2) key strategies in the Kansas Building an Environment for Science and Technology (B.E.S.T.) for Innovation report. In addition, this project contributes to improving Kansas cyberinfrastructure, with a focus on building a workforce that will improve the state economy through STEM education.The project tackles three barriers to the efficient and wide spread deployment of extreme-scale high-performance reconfigurable computing (HPRC) systems: (1) the imbalance in processing heterogeneity where processing resources are integrated in different ways and proportions, (2) lack of hierarchical parallel programming models and/or insufficiency of existing models, and (3) lack of accurate formal models that are capable of describing the instantaneous and transient behaviors of HPRCs rather than only describing average behaviors. These challenges, especially the first two, have hindered wide adoption of HPRC by a broad range of HPC users and more specifically scientists in domains such as quantum simulation, experimental High-Energy Physics (HEP), and Fossil Energy (FE). Three research activities are proposed in this project: (1) development of a portable node-level abstraction layer and run-time libraries that will manage and transparently provide to the end-user a unified view of the hardware resources independent of their type (virtualized hardware), while exploiting inherent features of FPGAs, (2) creation of a new parallel domain-specific language (DSL) derived from Partitioned Global Address Space (PGAS) models using a multi-layered C-to-hardware compilation that supports a system-wide multi-level hierarchy of granularity and parallelism while also providing application portability, and (3) formulation and implementation of formal models and discrete-event simulation tools for design space exploration based on stochastic Markov chains and queueing networks. The proposed concepts and techniques will be implemented and evaluated to demonstrate their portability on three distinct heterogeneous large-scale HPC systems: (1) an experimental multi-node HPC cluster populated with Xilinx FPGA accelerators, (2) an Intel Hardware Accelerator Research Program (HARP) system, and (3) a DS8 state-of-the-art OS-less HPRC system from DirectStream.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.