Active Projects

The goal of this project is to develop a partitioning framework for directed acyclic graphs (DAGs), that will result in balanced groups with minimal interconnecting edges, thereby facilitating the optimal scheduling of computational tasks encoded in the directed graphical structure. To this end, novel spectral graph methods will be developed, and current state-of-the-art multilevel partitioning algorithms will be adapted, taking into account the practical aspects and the specific requirements of task scheduling problems. The solid theoretical background of spectral methods, combined with the performance advantages that a multilevel scheme offers, will result in a partitioning software that surpasses other state-of-the-art algorithms in terms of the quality of the graph cuts, while maintaining competitive runtimes. The collaborating partners aim to bridge academic algorithmic research and industrial open-source software development, in order to design and implement the Multilevel Spectral - DAG partitioner (MS - DAG), that will guide the subsequent devising of a novel task scheduler.

This project aims to develop a scalable and efficient cloud-based method for sparse precision matrix estimation, a crucial task in the increasingly prevalent field of high-dimensional, low-sample-size (HDLSS) machine learning. If successful, this project will establish the foundation for standard or low-power systems to perform computationally demanding HDLSS machine learning and data analytics tasks common in many applications. Our project faces two key challenges: (C1) a hyperparameter tuning process that demands specialized domain expertise, and (C2) the high computational costs of solution methods. To overcome these, we require (R1) the elimination of the time-consuming process required for hyperparameter tuning, and (R2) an efficient and performant cloud-based solution that seamlessly integrates into existing workflows. Our objectives are (O1) to establish an algorithm that eliminates the need for hyperparameter tuning (i.e., tuning-free hyperparameters) and (O2) to develop a cloud-based solution with an API that leverages structural attributes of the computation for performance and scalability.

Computing has been disruptive to all scientific domains that increasingly rely on computational models and data to discover new knowledge and form decisions. With the explosion of Big Data, we are now faced with the ever-increasing size, variability, and uncertainty of the datasets. Some of the most challenging problems in data-driven science involve understanding the interactions between millions or even thousands of millions of variables. The vast quantity, veracity, velocity, and variety of data are challenging classical high-performance numerical methods and software for extreme-scale computing. Progress in research in scientific computing algorithms and software has been tightly linked to progress in microprocessor technology and high-performance programming technology. We are now in the process of embarking on the extreme-scale computing era which will revolutionize the architectural stack in a holistic fashion. It will also require research on optimized mathematical software libraries according to the device characteristics with novel numerical algorithms and data science applications that exploit them. How can we reconcile sustainable advances in sparse linear algebra and nonlinear optimization for new applications domains in data analytics while at the same time prepare for the anticipated sea-change looming in a twenty-year hardware horizon as well? We seek answers to these questions through computational methods that resolve fundamental challenges imposed by large-scale analytics, deep analysis, and precise predictions by advancing and preparing the foundation for the next generation of sparsified numerical methods. Our algorithms rely on the innovative coupling of sparsified numerical linear algebra and nonlinear optimization methods for data-intensive applications. The inherently deterministic character of these methods, when coupled with high communication demands, requires the development of robust approximation methods under the condition of extreme-scale computational science. This includes scientific libraries providing high-quality, reusable software components for constructing applications, as well as improved robustness ad portability. These developments will be driven by research on mathematical software, extreme-scale computing and an effort to push these developments toward applications. The focus on the computation of functions of matrix inverse entries presents a new dimension of numerical methods, since it goes beyond the classical requirements in solving linear systems or eigenvalue problems and has not yet been addressed in most of the research projects on massively parallel architectures. It is expected that the techniques developed by this will prove important in many of the other data-driven applications and will also provide basic tools for most of the applications for high performance computing (HPC) science and engineering problems. Novel, scalable software engineered to facilitate broader public use will be made available to the research and industrial community. Our numerical algorithms and mathematical software libraries are capable of leveraging emerging hardware paradigms and are applicable to a wide variety of existing applications such as finance, biology, health sciences, and many more. In particular, we will shed light on applications on nanoelectronic device simulation, and high-dimensional partial correlation estimations in genomics applications.

Advancing education and training in High Performance Computing (HPC) and its applicability to HPDA and AI is essential for strengthening the world-class European HPC ecosystem. It is of primary importance to ensure the digital transformation and the sustainability of high-priority economic sectors. Missing educated and skilled professionals in HPC/HPDA/AI could prevent Europe from creating socio-economic value with HPC. The HPC European Consortium Leading Education Activities aims to develop a new and innovative European Master programme focusing on high performance solutions to address these issues. The master programme aims at catalysing various aspects of the HPC ecosystem and its applications into different scientific and industrial domains. EUMaster4HPC brings together major players in HPC education in Europe and mobilizes them to unify existing programs into a common European curriculum. It leverages experience from various European countries and HPC communities to generate European added value beyond the potential of any single university. EUMaster4HPC emphasizes collaboration across Europe with innovative teaching paradigms including co-teaching and the cooperative development of new content relying on the best specialists in HPC education in Europe. Employers, researchers, HPC specialists, supercomputing centres, CoEs and technology providers will constitute a workforce towardEuroHPC projects this master in HPC pilot programme. This pilot will provide a base for further national and pan- European educational programmes in HPC all over Europe and our lessons learned and the material development will accelerate the uptake of HPC in academia and industry.The creation of a European network of HPC specialists will catalyse transfers and mutual support between students, teachers and industrial experts. A particular focus on mobility of students and teachers will enable students to rapidly gain experience through internships and exposure to European supercomputing centres.