A Learning-based Selection for Portfolio Scheduling of Scientific Applications on Heterogeneous Computing Systems

The execution of computationally intensive parallel applications in heterogeneous environments, where the quality and quantity of computing resources available to a single user continuously change, often leads to irregular behavior, in general due to variations of algorithmic and systemic nature. To improve the performance of scientific applications, loop scheduling algorithms are often employed for load balancing of their parallel loops. However, it is challenging to select the most robust scheduling algorithms that guarantee an optimized execution of scientific applications on large-scale computing systems. The computing systems comprise resources that are widely distributed, highly heterogeneous, often shared among multiple users, and have computing availabilities that cannot always be guaranteed or predicted. To address this challenge, in this work we focus on a portfolio-based approach to enable the dynamic selection and use of the most robust dynamic loop scheduling (DLS) algorithm from a portfolio of DLS algorithms, depending on the given application and current system characteristics, including the workload conditions. We provide a solution to the algorithm selection problem and experimentally evaluate its quality. We propose the use of supervised machine learning techniques to build empirical robustness prediction models that are used to predict a DLS algorithm's robustness for given scientific application characteristics and system availabilities. Using simulated scientific applications characteristics and system availabilities, along with empirical robustness prediction models, we show that the proposed portfolio-based approach enables the selection of the most robust DLS algorithm that satisfies a user-specified tolerance on the given application's performance degradation obtained in the particular computing system with a certain variable availability. We also show that the portfolio-based approach offers higher guarantees regarding the robust performance of the application using the automatically selected DLS algorithms when compared to the robust performance of the same application using a manually selected DLS algorithm. Keywords-Dynamic loop scheduling; Robustness; Algorithm selection; Empirical robustness prediction models; Machine learning techniques; Variable system availability