<div>As high performance computing systems are advancing from petascale to exascale, scientific workflows to integrate simulation and visualization/analysis are a key factor to influence scientific campaigns. As one of the campaigns to study fluid behaviors, computational fluid dynamics (CFD) simulations have progressed rapidly in the past several decades, and revolutionized our lives in many fields. Lattice Boltzmann method (LBM) is an evolving CFD approach to significantly reducing the complexity of the conventional CFD methods, and can simulate complex fluid flow phenomena with cheaper computational cost. This research focuses on accelerating the workflow of LBM simulation and data analysis.</div><div><br></div><div>I start my research on how to effectively integrate each component of a workflow at extreme scales. Firstly, we design an in-situ workflow benchmark that integrates seven state-of-the-art in-situ workflow systems with three synthetic applications, two real-world CFD applications, and corresponding data analysis. Then detailed performance analysis using visualized tracing shows that even the fastest existing workflow system still has 42% overhead. Then, I develop a novel minimized end-to-end workflow system, Zipper, which combines the fine-grain task parallelism of full asynchrony and pipelining. Meanwhile, I design a novel concurrent data transfer optimization method, which employs a multi-threaded work-stealing algorithm to transfer data using both channels of network and parallel file system. It significantly reduces the data transfer time by up to 32%, especially when the simulation application is stalled. Then investigation on the speedup using OmniPath network tools shows that the network congestion has been alleviated by up to 80%. At last, the scalability of the Zipper system has been verified by a performance model and various largescale workflow experiments on two HPC systems using up to 13,056 cores. Zipper is the fastest workflow system and outperforms the second-fastest by up to 2.2 times.</div><div><br></div><div>After minimizing the end-to-end time of the LBM workflow, I began to accelerate the memory-bound LBM algorithms. We first design novel parallel 2D memory-aware LBM algorithms. Then I extend to design 3D memory-aware LBM that combine features of single-copy distribution, single sweep, swap algorithm, prism traversal, and merging multiple temporal time steps. Strong scalability experiments on three HPC systems show that 2D and 3D memory-aware LBM algorithms outperform the existing fastest LBM by up to 4 times and 1.9 times, respectively. The speedup reasons are illustrated by theoretical algorithm analysis. Experimental roofline charts on modern CPU architectures show that memory-aware LBM algorithms can improve the arithmetic intensity (AI) of the fastest existing LBM by up to 4.6 times.</div>

The use of accelerators in heterogeneous systems is an established approach in designing petascale applications. Today, Compute Unified Device Architecture (CUDA) offers a rich programming interface for GPU accelerators but requires developers to incorporate several layers of parallelism on both the CPU and the GPU. From this increasing program complexity emerges the need for sophisticated performance tools. This work contributes by analyzing hybrid MPICUDA programs for properties based on wait states, such as the critical path, a metric proven to identify application bottlenecks effectively. We developed a tool to construct a dependency graph based on an execution trace and the inherent dependencies of the programming models CUDA and Message Passing Interface (MPI). Thereafter, it detects wait states and attributes blame to responsible activities. Together with the property of being on the critical path, we can identify activities that are most viable for optimization. To evaluate the global impact of optimizations to critical activities, we predict the program execution using a graph-based performance projection. The developed approach has been demonstrated with suitable examples to be both scalable and correct. Furthermore, we establish a new categorization of CUDA inefficiency patterns ensuing from the dependencies between CUDA activities.