Date of Award
Doctor of Philosophy (PhD)
School of Computing
Jon C. Calhoun
Heterogeneous systems with both CPUs and GPUs have become important system architectures in emergent High Performance Computing (HPC) systems. Heterogeneous systems must address both performance-scalability and power-scalability in the presence of failures. Aggressive power reduction pushes hardware to its operating limit and increases the failure rate. Resilience allows programs to progress when subjected to faults and is an integral component of large-scale systems, but incurs significant time and energy overhead. The future exascale systems are expected to have higher power consumption with higher fault rates. Sparse data computation is the fundamental kernel in many scientific applications. It is suitable for the studies of scalability and resilience on heterogeneous systems due to its computational characteristics.
To deliver the promised performance within the given power budget, heterogeneous computing mandates a deep understanding of the interplay between scalability and resilience. Managing scalability and resilience is challenging in heterogeneous systems, due to the heterogeneous compute capability, power consumption, and varying failure rates between CPUs and GPUs. Scalability and resilience have been traditionally studied in isolation, and optimizing one typically detrimentally impacts the other. While prior works have been proved successful in optimizing scalability and resilience on CPU-based homogeneous systems, simply extending current approaches to heterogeneous systems results in suboptimal performance-scalability and/or power-scalability.
To address the above multiple research challenges, we propose novel resilience and energy-efficiency technologies to optimize scalability and resilience for sparse data computation on heterogeneous systems with CPUs and GPUs. First, we present generalized analytical and experimental methods to analyze and quantify the time and energy costs of various recovery schemes, and develop and prototype performance optimization and power management strategies to improve scalability for sparse linear solvers. Our results quantitatively reveal that each resilience scheme has its own advantages depending on the fault rate, system size, and power budget, and the forward recovery can further benefit from our performance and power optimizations for large-scale computing. Second, we design a novel resilience technique that relaxes the requirement of synchronization and identicalness for processes, and allows them to run in heterogeneous resources with power reduction. Our results show a significant reduction in energy for unmodified programs in various fault situations compared to exact replication techniques. Third, we propose a novel distributed sparse tensor decomposition that utilizes an asynchronous RDMA-based approach with OpenSHMEM to improve scalability on large-scale systems and prove that our method works well in heterogeneous systems. Our results show our irregularity-aware workload partition and balanced-asynchronous algorithms are scalable and outperform the state-of-the-art distributed implementations. We demonstrate that understanding different bottlenecks for various types of tensors plays critical roles in improving scalability.
Miao, Zheng, "Scalable and Reliable Sparse Data Computation on Emergent High Performance Computing Systems" (2022). All Dissertations. 2972.