UC Santa Cruz

Easy access to powerful computational resources (cloud platforms and supercomputers) has revolutionized scientific research in the 21st century. However, these resources are often under-utilized, resulting in sub-optimal performance and increased operational costs. But, improving application performance is a complex task, as various factors such as mismatch between hardware configuration, runtime environment, and source code can contribute to sub-optimal performance.

To address these challenges, our study describes a `profile-and-optimize' approach that leverages profiling data obtained by running applications on supercomputing systems. By systemically analyzing this data and modifying both the runtime environment and source code, we aim to improve application performance and scalability. We apply our methodology to four case studies and showcase performance improvements by comparing baseline results with optimized results obtained through our approach. Through our methods and the insights gained from the case studies, we provide a valuable resource for optimizing similar applications. By following our process, researchers and developers can reduce application runtime, maximize the utilization of computational resources, and run finer-grained simulations.