Date of Award
Doctor of Philosophy
Jack Dongarra, George Bosilca, Michael Jantz, Hatem Ltaief
The last decades have witnessed a rapid improvement of computational capabilities in high-performance computing (HPC) platforms thanks to hardware technology scaling. HPC architectures benefit from mainstream advances on the hardware with many-core systems, deep hierarchical memory subsystem, non-uniform memory access, and an ever-increasing gap between computational power and memory bandwidth. This has necessitated continuous adaptations across the software stack to maintain high hardware utilization. In this HPC landscape of potentially million-way parallelism, task-based programming models associated with dynamic runtime systems are becoming more popular, which fosters developers’ productivity at extreme scale by abstracting the underlying hardware complexity.
In this context, this dissertation highlights how a software bundle powered by a task-based programming model can address the heterogeneous workloads engendered by HPC applications., i.e., data redistribution, geospatial modeling and 3D unstructured mesh deformation here. Data redistribution aims to reshuffle data to optimize some objective for an algorithm, whose objective can be multi-dimensional, such as improving computational load balance or decreasing communication volume or cost, with the ultimate goal of increasing the efficiency and therefore reducing the time-to-solution for the algorithm. Geostatistical modeling, one of the prime motivating applications for exascale computing, is a technique for predicting desired quantities from geographically distributed data, based on statistical models and optimization of parameters. Meshing the deformable contour of moving 3D bodies is an expensive operation that can cause huge computational challenges in fluid-structure interaction (FSI) applications. Therefore, in this dissertation, Redistribute-PaRSEC, ExaGeoStat-PaRSEC and HiCMA-PaRSEC are proposed to efficiently tackle these HPC applications respectively at extreme scale, and they are evaluated on multiple HPC clusters, including AMD-based, Intel-based, Arm-based CPU systems and IBM-based multi-GPU system. This multidisciplinary work emphasizes the need for runtime systems to go beyond their primary responsibility of task scheduling on massively parallel hardware system for servicing the next-generation scientific applications.
Cao, Qinglei, "Task-based Runtime Optimizations Towards High Performance Computing Applications. " PhD diss., University of Tennessee, 2022.