Enhancing Performance in Scientific Applications with Energy and Resilience Constraints from Modern Architectures

Abstract

An overarching goal in the scientific computing field is maximizing performance; however, on modern architectures, performance can be constrained by energy efficiency and resilience. Enhancing these attributes, however, increasingly becomes a difficult task. Despite a significant body of work that has been developed in this field to enhance these attributes, implementing a high-performing solution is still often difficult due to complex computation patterns and environmental constraints, i.e. those imposed by hardware, software, and run-time configurations. This dissertation provides a feasibility study for a framework which suggests optimizations that are likely to increase an arbitrary scientific application's performance, energy efficiency, and resilience. The scope of this study is limited to multicore architectures and three aspects of scientific applications: computation-, storage-, and domain-level Within the context of these three aspects, we address four key challenges in optimizing scientific applications: 1) better managing a system's hardware resources among co-running processes and threads, 2) implementing fault tolerance while maintaining a certain degree of performance, 3) identifying the most profitable optimizations for performance based on domain-level properties and environmental configurations, and 4) maintaining program correctness in optimizing applications. To address these challenges, we present a new model, the tri-dimensional application model, which models the computation-, storage-, and domain-level aspects of scientific applications, in order to identify opportunities for potential optimization. We present a framework that uses this model to identify and apply optimizations that are likely to benefit performance, energy efficiency, or resilience, along with a feasibility study through four components. First, we model computation-level information in order to better manage a system's hardware resources among co-running processes and threads, where our results lead to an average increase in the energy efficiency of a system by 12%. Second, we model storage-level information in order to implement fault tolerance while maintaining a certain degree of performance. This component focuses on using NVRAM to support faster resilience and instant recovery of individual nodes in a multi-node environment. Third, we model domain-level information in order to identify the most profitable optimizations for performance based on domain-level properties and environmental configurations. This component uses a prediction model to create optimization strategies, which we show outperform an optimized grid computing library by 2.8x on average. Fourth, we apply compiler optimizations as source-level transformations through an integrated program development environment for developers to conveniently improve the efficiency of their applications. Our optimizations outperform the ICC and GCC compilers with the -O3 optimization flag by 1.84x and 3.83x on average. To showcase our developed workflow, we study ocean_cp from Splash 2, where we increase its performance by 6%, energy efficiency by 4%, and implement a resilience strategy that leverages NVRAM that performs only 6% worse than the theoretical fastest strategy.