Articles

High-Performance Computing (HPC) is indispensable for large-scale scientific simulations, but achieving optimal performance on modern supercomputers is increasingly challenging. As HPC systems scale toward exascale, they face escalating complexity in hardware, software, and workloads. Traditional optimization methods (manual tuning and heuristic algorithms) struggle to cope with dynamic workloads and intricate system behaviors. Artificial Intelligence (AI) techniques offer a promising approach to address these challenges. This article provides an overview of AI-powered optimization in HPC, focusing on how machine learning and related AI methods enhance performance, efficiency, and scalability of scientific simulations. We survey key AI techniques applied to HPC optimization including machine learning for performance modeling, deep reinforcement learning for resource management, and AI-driven surrogate models for accelerating simulations and illustrate their impact through case studies in domains such as job scheduling and fluid dynamics. We discuss the practical applications of these techniques, highlighting reported performance gains (e.g., substantial reductions in simulation run time and improved resource utilization). We also examine the challenges in integrating AI with HPC (such as training overhead, data movement, and reliability concerns) and outline future directions for research. The convergence of AI and HPC is poised to produce “smart” simulation workflows that intelligently adapt and optimize in real time, pushing the frontiers of scientific computing.

To ensure maximum flexibility, service providers offer a variety of computing options with regard to CPU, memory capacity, and network bandwidth. At the same time, the efficient operation of current cloud applications requires an infrastructure that can adjust its configuration continuously across multiple dimensions, which are generally not statically predefined. Our research shows that these requirements are hardly met with today's typical public cloud and management approaches. To provide such a highly dynamic and flexible execution environment, we propose the application-driven autonomic management of data center resources as the core vision for the development of a future cloud infrastructure. As part of this vision and the required gradual progress toward it, we present the concept of composable infrastructure and its impact on resource allocation for multi-cloud environments. We introduce relevant techniques for optimizing resource allocation strategies and indicate future research opportunities [1]. Many cloud service providers offer computing instances that can be configured with arbitrary capacity, depending on the availability of certain hardware resources. This level of configurability provides customers with the desired flexibility for executing their applications. Because of the large number of such prerequisite instances with often varying characteristics, service consumers must invest considerable effort to set up or reconfigure elaborate resource provisioning systems. Most importantly, they must differentiate the loads to be distributed between jobs that need to be executed versus placeholder jobs, i.e., jobs that trigger the automatic elasticity functionality responsible for resource allocator reconfiguration. Operations research reveals that the optimization of resource allocator reconfiguration strategies is a fundamentally difficult problem due to its NP-hardness. Despite these challenges, dynamic resource allocation in multi-clouds is becoming increasingly important since modern Internet-based service settings are dispersed across multiple providers [2].

Green cloud computing is part of endeavors to develop sustainable data center ecosystems and, more importantly, nurtures a mindful alignment between environmental considerations and our cloud computing practices. This view is reinforced with the requirements of resource and energy minimization, as well as clean computing. This paper surveys the current practices, strategies, and significant aspects involved in moving towards green cloud computing, providing energy-efficient data centers. The energy efficiency criteria call for unified strategies in power-proportional components, big data storage, server systems, and power supply units to save holistic energy. In addition, there are significant challenges in moving towards green cloud computing for service providers and data center operators. We address various energy-conscious resource management technologies and discuss the importance of developing innovative, effective green management solutions. Data centers are ubiquitous but inherently more conspicuous to begin to see the urgency of making them sustainable in our ecological environment. With this in mind, this paper encapsulates the multidimensional issues and increased complexities of bringing up green solutions in cloud computing practices and provides guidance and potential strategies. We outline, realign, and insist on adopting strategies in practice not only from the technical aspect but also in strengthening partnerships and investigating strategies to further dissect challenges, converge solutions, and consider our impact in even more areas of study.