The European master for HPC curriculum

Contents lists available at ScienceDirect Journal of Parallel and Distributed Computing journal homepage: www.elsevier.com/locate/jpdc The European master for HPC curriculum Pascal Bouvrya, Mats Brorsson a, Ramon Canalb, Aryan Eftekhari c, Siegfried Höfingerd, Didier Smets e, Harald Köstler f, Tomáš Kozubekg, Ezhilmathi Krishnasamy a,, Josep Llosa b, Alexandra Lukas-Rotherf, Xavier Martorellb, Dirk Pleiter h, ,∗, Ana Proykova i, Maria-Ribera Sancho b, Olaf Schenkc, Cristina Silvanoj aParallel Computing & Optimisation Group and FSTM-DCS/SnT, University of Luxembourg, Esch-sur-Alzette, 4364, Luxembourg bUniversitat Politècnica de Catalunya, Barcelona, 08034, Spain cInstitute of Computing, Faculty of Informatics, Università della Svizzera italiana, Lugano, 6962, Switzerland dTechnische Universität Wien, Wien, 1040, Austria eLaboratoire Jacques-Louis Lions, Sorbonne Université, Paris, 75005, France fDepartment of Computer Science, Friedrich-Alexander-Universität Erlangen-Nürnberg, Erlangen, 91058, Germany gIT4Innovations, VSB-Technical University of Ostrava, Ostrava, 70800, Czech Republic hDivision of Computational Science and Technology, KTH Royal Institute of Technology, Stockholm, 10044, Sweden iHPC laboratory, Faculty of Physics, University of Sofia, Sofia, 1164, Bulgaria jDipartimento di Elettronica, Informazione e Bioingegneria, Politecnico Milano, Milano, 20133, Italy A R T I C L E I N F O A B S T R A C T Keywords: Computing education High-performance computing Master in HPC Model curricula The use of High-Performance Computing (HPC) is crucial for addressing various grand challenges. While significant investments are made in digital infrastructures that comprise HPC resources, its realisation, operation, and, in particular, its use critically depends on suitably trained experts. In this paper, we present the results of an effort to design and implement a pan-European reference curriculum for a master’s degree in HPC. 1. Introduction The use of High-Performance Computing (HPC) is critical to address many of the grand scientific, engineering, and societal challenges. The scientific challenges requiring HPC range from solving fundamental open questions of physics at the largest and smallest scales to challenges in life sciences like the decoding of the human brain. One of the most prominent challenges where HPC plays a crucial role is climate adaptation. Model simulations using HPC systems have played a key role in predicting climate change. Destination Earth is currently developing large-scale climate and weather digital twins, which will push the use of HPC to the next level [20]. The importance of HPC is also reflected in the growing volume of the market. Analysts from Hyperion Research *Corresponding author. E-mail address: [email protected] (D. Pleiter). 1https://eurohpc-ju.europa.eu/. predict the annual growth of the HPC market of 8.2 % over the next five years [6]. The role of HPC has also been recognized at the political level. For example, in Europe, a new organisation has been created called Joint Undertaking EuroHPC,1which establishes an exascale HPC infrastructure and supports related research. The strategy is to coordinate efforts and pool resources in order to facilitate a globally leading role of European stakeholders. Similar efforts are ongoing in other regions of the world, most notably in China, India, Japan, and the US. The success of these efforts critically depends on the availability of appropriately trained researchers, engineers, and other experts. Ensuring the availability of such experts is challenging due to the breadth of required skills coming from various science and engineering domains. HPC is a very interdisciplinary area, which needs to be reflected in the https://doi.org/10.1016/j.jpdc.2025.105081 Received 1 June 2024; Received in revised form 22 December 2024; Accepted 1 April 2025 Journal of Parallel and Distributed Computing 201 (2025) 105081 Available online 3 April 2025 0743-7315/© 2025 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY license ( http://creativecommons.org/licenses/by/4.0/ ). P. Bouvry, M. Brorsson, R. Canal et al. education. This ranges from various domains of computational science and engineering, from computer science to computer and electrical engineering. To address the challenge of the future workforce, EuroHPC has provided funding for the EUMaster4HPC project.2The project consortium includes dozens of universities and educational institutions throughout Europe, HPC centres, and private companies that are active in the area of HPC. A primary objective of the project is the design of a common European master reference curriculum in HPC. It created an opportunity to formulate a holistic curriculum that draws on the knowledge and experience of a wide range of stakeholders, expands existing curricula with unique characteristics, and promotes new teaching and training methods. Examples are the training of transversal skills or the implementation of challenges. The curriculum is designed in a modular way to facilitate its uptake by a large number of universities throughout Europe. This makes it easier to implement it within existing programs. Furthermore, the project implements a pilot program involving about 150 students, who benefit from a mobility program to receive the most advanced education and training from multiple organisations. With this paper, we make the following contributions: 1. Documentation of a holistic curriculum for a master’s in HPC education, with a focus on its unique features. 2. Discussion of the methodology that led to the formulation of this curriculum. 3. Evaluation of early experiences with the implementation of the curriculum and the training of EUMaster4HPC students. This paper is organised as follows: In Section 2, we discuss a selection of related work. The next Section 3, documents the methodology that resulted in the curriculum, which is documented in Section 4. In Section 5, we report on different aspects of the initial implementation of the curriculum. In Sections 6and 7, we provide a first evaluation of the use of the curriculum based on a case study and a survey. We end with a summary and conclusions in Section 8. 2. Related work The importance of quality education in HPC for basic sciences and engineering has long been recognised (see, for example, the early report by Long et al. [7]). An interesting newer development are the offerings to students who do not major in computer science, but instead are introduced to core concepts, thereby stressing the new objective of computational thinking [21]. EUMaster4HPC has extended the scope further by also considering HPC system administrators as a possible future role, which is neither covered by education for system administration nor the traditional HPC education. Such an extension has earlier been proposed in [24]. One particular challenge for education in the area of HPC is the fast evolution of the computational methods as well as the architectures and technologies. One notable example is the introduction of heterogeneous computing platforms. Qasem et al. suggested addressing the growing need for handling such platforms by defining modules that introduce fundamental concepts of heterogeneous (or accelerated) computing into undergraduate education in computer science and computational engineering [13]. Another strategy is the stronger involvement of supercomputing centres. A practical example is given by Chen et al. [3] reporting on efforts at China’s National University of Defense Technology (NUDT) that has hosted the Tianhe supercomputers. Over the years, numerous studies have been published introducing new methods for teaching HPC. One example is the organisation of hackathons as part of university education, of which the Brazilian 2https://eumaster4hpc.uni.lu/. Parallel Programming Marathons [8] is an example. The authors consider hackathons as a good way of helping students with a paradigm shift from traditionally sequential programming techniques to parallel and distributed computing. It also relates to our efforts to involve students from different universities in obtaining practical experiences on HPC systems (see Section 5.4.3), which others have proposed as an educational tool (see, e.g., [5]). Another example of papers describing the successful adoption of new ways of teaching HPC is the use of Jupyter notebooks [9]. Building small-scale, low-cost computing clusters based on simple devices, e.g. edge devices like Raspberry Pis, is not new (see, e.g., [4]) but still has a limited uptake. The integration of such a methodology within courses is discussed in [19]. More recently, Xu Zhiguang combined the approach of building clusters on-site using low-cost devices with the use of high-end computing resources in the cloud leveraging Google collabs [23]. Various works document the development of specific courses. One nice recent example is Varbanescu et al. [18] who developed a full performance engineering courses and documented the setup. Such courses fill an important gap as there is a need for methodology education in the area of performance analysis and engineering [1]. EUMaster4HPC addresses this by foreseeing a specialisation as Performance Analyst and Advisor (see Section 4.2.2). A broader scope initiative in curriculum development for Parallel and Distributed Computing has been launched with the guidelines published by the NSF/IEEE-TCPP Curriculum Working Group [11]. An analysis of current trends in HPC education and corresponding challenges for future directions has been reported by the Working Group on Innovation and Technology in Computer Science Education [14]. Competencybased learning was identified as a needed transformation from the more prevalent knowledge-based learning in HPC education [2]. A good and continuous source of publications and presentations related to teaching methods and new courses in the area of HPC education are several workshop series: •The EduPar workshops take usually place as part of the IEEE International Parallel & Distributed Processing Symposia (IPDPS) [10], •EduHPC is a series of workshops in the context of the International Conferences for High Performance Computing (SC).3 •Edu-HiPC workshops have been organised for a few years on an annual basis in India.4 In the context of EuroHPC, a related effort is to define a training baseline that is not limited to university education at the master’s level. This effort is being carried out by the CASTIEL2 and EuroCC2 projects.5The training baseline for HPC, High-Performance Data Analytics (HPDA), and Artificial Intelligence (AI) is defined around six different career path profiles. One of the main design goals of the training baseline was that each of the building blocks remain modular, allowing for further adaptations in future releases. The target sectors comprise “Research/Academic + + ”, “SMEs” and “Big Industry” and corresponding career path profiles are “HPC application developer”, “System Administrator”, “HPC application user”, “Data scientist / Analyst”, “Executive / Technical lead” and “System designer”. Each career path profile is assigned a specific list of training units, providing a stepwise acquisition of skills needed in the respective target profession. A notable difference to the EUMaster4HPC curriculum is the significantly reduced workload for a particular career path profile. Career path profiles are considered more as additional training within the realm of lifelong learning [15] to offer interested candidates a non-traditional 3For the 2024 edition, see https://tcpp.cs.gsu.edu/curriculum/?q= eduhpc24. 4For the 2024 edition, see https://tcpp.cs.gsu.edu/curriculum/?q= eduHiPC24. 5https://www.eurocc-access.eu/. Journal of Parallel and Distributed Computing 201 (2025) 105081 2 P. Bouvry, M. Brorsson, R. Canal et al. route to specialization. Preliminary discussions had identified microcredentials [16] as a viable way of certification of the successful completion of a particular career path profile — albeit at the lower end of approximately 3-6 ECTS. Given the explicit emphasis on industry, the training baseline could offer interesting extensions to the EUMaster4HPC curriculum. 3. Methodology The process of developing the curriculum has been a multistage effort. It started with an analysis of the need for experts trained in HPCrelated topics in the private and academic sectors. This resulted in a set of skill profiles. Initial results of this effort have been published earlier [12]. For the design of the curriculum, a balance had to be maintained between fixed characteristics and variable elements. The fixed characteristics include a duration of 4 semesters. To obtain a master’s degree, students are expected to take courses that translate into 30 ECTS per semester and a total of 120 ECTS to obtain a master’s degree. To keep the curriculum flexible, the teaching content has been defined in terms of small modules with 1 to 3 ECTS per module and an effort of approximately 10 to 30 teaching hours. Although the set of modules, presented in the next Section, is based on current practices, it has been augmented and adjusted to take into account new trends and upcoming needs into account. To assess the upcoming private sector requirements, a dialogue has been conducted with relevant companies through questionnaires, structured interviews, and discussions during conferences and other meetings related to HPC. The interaction was guided by an analysis of job advertisements. As these reflect only current needs, the dialogue focused on future trends. The result of this effort has been the identification of six different skill profiles: •Computational scientist, HPC application support specialist, •Application software developer, •HPC system software developer, HPC software engineer, •HPC hardware developer, •HPC system administrator, and •HPC architect. A similar approach has also been taken for the academic sector. The results have been the following four profiles: •Applications •Parallel programming and tools support •DevOps (system support and development) • System architect The assessment of the skill requirements in both sectors also confirmed the need to offer education and training in transversal and professional skills. These skills are often not yet addressed in existing curricula for HPC studies. We considered a wide range of skills, including those related to research techniques, topics related to entrepreneurship, and skills to address legal topics. For developing this part of the curriculum, the following steps have been performed: •Analysis of existing curricula and initiatives across the organisations involved in the EUMaster4HPC project. •Selection of relevant transversal skills and respective learning outcomes. •Exploration of the educational approaches to achieve the expected learning outcomes. For an overview of this part of the curriculum, see Section 4.3. For each of the modules, learning outcomes have been defined, taking Bloom’s taxonomy into account. The main strength of our perspective is the adoption of a structural methodology to offer a comprehensive Fig. 1. Proposed fundamental modules and curriculum concept in EUMaster4HPC. approach to curriculum design. This methodology enables university programs to consider student requirements, evaluate the feasibility of teaching materials and pedagogical approaches, and develop assessments with concrete learning objectives. Bloom’s taxonomy provides a framework for understanding the learner’s needs, which is useful at the graduate level, as students may be accepted for reasons other than ability. A new type of module, balancing modules, has been introduced to offer more flexibility in terms of study paths. These modules support the intake of students with bachelor’s degrees other than computer science. An overview of the proposed balancing modules is provided in Section 4.4. 4. The curriculum In this section, the curriculum is introduced. A visual overview is provided in Fig. 1It shows the different domains of the fundamentals, which are introduced in Sections 4.1 and 4.3. The specialisations will be discussed in Section 4.2. Finally, the balancing modules will be presented in Section 4.4. 4.1. Fundamentals We have identified five domains or specific areas for the fundamentals: •Mathematics and Statistics •Software Engineering •Parallel Programming •Computer Architecture •Transversal Skills The approach that we followed was to define small modules that can be combined to provide a number of courses, see Fig. 2. Transversal skills address issues like diversity, inclusion, and gender bias as a cross-cutting topic for all modules. Modules are set to have 3 ECTS so that by combining them it is easy to build regular courses of 6 ECTS, or larger ones, up to 9 ECTS. Each university imparting the master can decide to combine the modules as they prefer, provided that the final content is comparable. Some of the modules in each domain are “basic”, in the sense that they overlap with contents potentially done in a previous bachelor’s degree. Those modules are only assigned to students with different backgrounds. For example, students coming from a nontechnical bachelor can be assigned modules such as “Basic Computer Architecture”, “Basic Parallel Programming”, or “Operating Systems”. Table 1shows the weight of each area in the curriculum, as well as the amount of knowledge implemented as Balancing Modules, and the amount contributing to the core of Fundamentals. The following subsections show the organization of the modules for each domain and their learning outcomes: Journal of Parallel and Distributed Computing 201 (2025) 105081 3 P. Bouvry, M. Brorsson, R. Canal et al. Fig. 2. Proposed fundamentals in EUMaster4HPC (Year One). Table 1 Proposed ECTS for fundamentals and balancing modules in year one. Area Balancing ECTS Fundamentals ECTS Total Mathematics and Statistics 6 6 12 Software Engineering 6 – 6 Parallel Programming 9 24 33 Computer Architecture 3 15 18 Transversal Skills – 9 9 Total 24 54 78 4.1.1. Mathematics and statistics This domain gets the students to a common level, by allowing those with no or low exposure to Algebra or Statistics to take the Balancing Modules of Numerical Methods for Linear Systems and Introduction to Statistics and Learning. Once all students have been brought to a common level, the master program includes Numerical Methods for Partial Differential Equations, including the application of numerical analysis, considering stability, consistency, and rate of convergence, and the methods of finite elements, Galerkin, Fourier and spectral methods. On the practical side, students develop and compare the performance of different numerical methods. Later in the domain, Numerical Optimization includes the knowledge of mathematical aspects of optimization, smooth and non-smooth optimization techniques, stochastic gradient and Limited-memory BroydenFletcher-Goldfarb-Shanno (L-BFGS) algorithms. On the practical side, students develop sample applications using those optimization methods. 4.1.2. Software engineering Software engineering techniques and tools are already common for students coming from CS-related backgrounds. For this reason, we consider these topics to belong to Balancing Modules. Two modules are defined: Software Development Techniques and Tools including the description and use of the software life cycle models, standard procedures (like ISO/IEC/IEEE 24765), requirements, use-cases, portability and performance portability; Software Quality, Validation and Testing includes the consideration and implementation of software sustainability, technical depth assessment, software quality criteria, testing methods, and their practical application. 4.1.3. Parallel programming The domain of Parallel Programming includes 3 modules from the set of Balancing topics that are usually overlapping with CS-related bachelors: Operating Systems, Basic Parallel Programming, and Algorithms and Data Structures. On top of those, the master curriculum provides 4 additional areas of knowledge: Parallel programming tools The area includes Compilers and Optimizations, and Performance Analysis and Models. In this area the students get to know, the abilities to explain and use the concepts related to the compilation process, programming languages, code optimizations, exploitation of the memory hierarchy, main techniques for obtaining parallelism (instruction-level parallelism, vectorization, and multi-threading), performance analysis metrics and laws (Amdahl’s and Gustafson’s law), performance analysis models, and tools. On the more practical side, use profiling-based optimizations, and detect overhead and bottlenecks on parallel applications. Shared-memory environments and programming The area covers the description and use of single-node shared-memory computers, applying parallelization techniques using well-known, standard, and state-of-theart parallel programming models (currently proposed are MPI, OpenMP, SYCL, etc.). Use of automatic parallelization, and combine parallelization and vectorization. On the practical side, apply parallel constructs, tasking, and vectorization to particular benchmarks and code structures. Heterogeneous environments and programming The area includes the description and use of single-node shared-memory systems with accelerators (currently GPUs, TPUs, DPUs and FPGAs), applying offloading techniques and managing data transfers with standard and state-of-the-art heterogeneous programming tools (currently CUDA, OpenCL, OpenACC, OpenMP target, SYCL, etc.), improve the use of the accelerators for particular algorithms. On the practical side, be able to analyse the performance of heterogeneous parallel applications and detect drawbacks. Distributed-memory environments and programming The area covers the description and use of cluster, grid and cloud environments, compilation tools, organisation of applications based on workflows and pipelines, microservices, and parallelization of applications with standard, state-ofthe-art distributed programming tools (currently MPI, etc.). On the practical side, use the techniques to exploit interoperability across different programming models (currently MPI + OpenMP, MPI + CUDA, etc.). 4.1.4. Computer architecture The domain of Computer Architecture includes one module from the set of Balancing topics that usually overlaps with CS-related bachelors: Basic Computer Architecture. Journal of Parallel and Distributed Computing 201 (2025) 105081 4 P. Bouvry, M. Brorsson, R. Canal et al. On top of those, the master curriculum provides 4 additional areas of knowledge: Processor architecture The area covers basic and advanced processor architecture. In this area, the students get to know the different types of processor architectures, the fundamental mechanisms for static ILP processors, software-level instruction scheduling techniques (list scheduling, unrolling, software pipelining), the memory hierarchy and its components, cache design and cache strategies and protection and security issues related to the processor architecture. On the more advanced side, students learn and get to use the fundamental mechanisms for dynamic ILP processors (out-of-order execution, branch prediction, exception handling, instruction and data prefetch, etc.) and consider protection and security issues in the processor design. On the practical side, analyse the performance and assess the impact on power and energy of singlethreaded programs on in-order and out-of-order processors, and apply simple optimizations to code fragments. Heterogeneous computer architecture The area covers accelerators (e.g., GPUs, FPGAs), their architecture and micro-architecture, hardware techniques for thread scheduling and synchronization, accelerators memory hierarchy, FPGA building blocks and FPGA interconnects. On the practical side, explore accelerator performance capabilities for synthetic benchmarks, program sample applications with FPGA programming workflows and explore the performance of selected computational motifs on accelerators (e.g., structured grids, FFT, dense linear algebra). Multiprocessor architecture The area covers the implementation of different models of thread-level and data-level parallelism, such as core multi-threading, many-core processors, vector units, multi-chip modules, the impact of multiprocessing on the processor ISAs, implementation of hardware-level synchronization mechanisms, the operation of the memory hierarchy on multiprocessors: snooping and directory-based coherence protocols, coherency, and consistency. Design at the architectural level of hardware for shared-memory, basic components of modern parallel systems, including multiple processors, cache hierarchies and networks. And quantify performance metrics for parallel systems and evaluate parallel computer organizations. On the practical side, develop small benchmarks evaluating the performance of hardware synchronization mechanisms, the overhead of thread and process management and evaluate the mechanisms implementing data coherence and consistency. Networking The area covers network topologies, routing algorithms, flow control, congestion control mechanisms, error detection/correction mechanisms, ISO/OSI model, network protocols (TCP/IP, RDMA, etc.) and network performance metrics (e.g., latency, bandwidth, and quality of service). On the practical side, develop benchmarks implementing communications with commonly used protocols (TCP/IP, RDMA, etc.), and analyse the performance of a network considering latency, bandwidth and quality of service. 4.1.5. Transversal skills The Transversal Skills are distributed across the different modules of the curriculum. In this paper, we collected all the information, and it is presented in Section 4.3. 4.2. Specialisations Specialisations are typically offered in the 3rd semester of the Master’s degree. Within the EUMaster4HPC project, we have identified the following four specialisations as the most significant through the analysis of the job market around the private and academic sectors: •Numerical and Data Specialist for Science Domains •Performance Analyst and Advisor • System Development and Support Table 2 Proposed ECTS for specialisations and modules in year two. Specialisation Modules ECTS Numerical and Data Specialist for Science Domains 20 66 Performance Analyst and Advisor 15 51 System Development and Support 17 57 System Architect 16 54 • System Architect For each specialisation, several teaching modules have been identified, including all relevant and actual topics in the form of learning outcomes that will allow students to gather additional knowledge and expertise within the selected specialisation. Each module has its prerequisites, the contents the students will learn, and various suggested practical topics. The introduced specialisations complement the fundamentals presented above and together are the basis for building the general curriculum of the master program. The number of modules currently identified and respective ECTS offered by each specialisation are shown in Table 2. Each specialisation should contain a minimum of 30 ECTS of modules. Practically, they have more ECTS to allow more freedom to the implementing institutions. The number of ECTS to pass specialisation is 30. Each specialisation contains regular modules for 3 ECTS, a project module for a maximum of 9 ECTS, and a list of modules shared with other specialisations. All modules are grouped into thematic areas (see Fig. 3), which provide the students with the learning outcomes identified for each specialisation and further described in the following subsections. The availability of courses in different implementations at different institutions may result in slight changes in the modules and ECTS. 4.2.1. Numerical and data specialist for science domains This specialisation focuses on graduate experts in scientific computing, mathematical modelling, the development of numerical algorithms for supercomputers, parallel programming, and knowledge of some application domains, such as engineering, physics, and biomedical. Furthermore, AI modelling with HPC capability and high-performance data analytics that are available for targeting research and applications is trending in AI / ML. They will be able to interact in interdisciplinary teams with experts and scientists in the application domain on the development of efficient parallel implementations, performance analysis, large data analysis, and visualisation. The following will discuss more details about the learning outcomes of this specialisation. Fig. 4and Table 3show an overview of the areas targeted and the expected ECTS within this specialisation. A detailed description of each area follows. Scientific computing The topic will focus on efficient numerical algorithms and analysis for differential equations, especially for partial differential equations. In particular, this area contains modules such as Computational Numerical Linear Algebra, Parallel Numerical Methods, and Implementation of Partial Differential Equation (PDE) Solvers. They include algorithms for linear algebra: orthogonalisation, least-squares methods, solving large linear systems, and studying their stability and communication requirements. For example, how good is the stability of the algorithm with respect to the level of parallelism? Furthermore, different domain decomposition methods will be studied in the context of parallel numerical methods, such as multigrid and Schwarz methods. Finally, we introduce different well-known numerical methods for solving PDE that will be studied in detail, such as finite difference, finite volume and finite element methods. The study group is expected to implement any of those methods from scratch, and also to have hands-on experience with handling (or using) libraries used within those numerical methods. Importantly, all of those methods and techniques will always be discussed in the context of parallel architecture machines, including GPU-accelerated systems. Journal of Parallel and Distributed Computing 201 (2025) 105081 5 P. Bouvry, M. Brorsson, R. Canal et al. Fig. 3. Proposed specialisations in EUMaster4HPC (Year Two). Fig. 4. Proposed modules (study areas) within Numerical and Data Specialists for Science Domains. Advanced HPC topics for science domains It introduces advanced topics such as quantum computing and domain-specific programming models. Regarding quantum computing, it covers quantum computer architecture (for example, qubits) and programming models (for example, Qiskit, Ocean, Cirq, etc.) and applying the quantum computing technique (using the real machine or simulator) to solve applications from science and engineering. Furthermore, within domain-specific programming models, especially focusing on embedded systems and their programming models. For example, programmability and running numerical or AI algorithms such as in FPGAs and IPU architectures. Finally, it introduces how to run a large science and engineering simulation (for example, medical engineering, fluid dynamics, solid mechanics and dynamics, life sciences, material sciences, and chemistry) using large HPC machines, in particular, by studying software packages for simulation, parallelisation strategies, methods, mathematical properties, and implementation approaches on HPC systems. AI/ML It covers the essential topics of expertise in machine learning, artificial intelligence, and computational statistics. Machine learning will focus on topics such as model evaluation and selection, anomaly detection, conformal learning (prediction with guarantees of accuracy), causal inference (identification of causal relationships), and hybrid approaches such as integrating numerical methods with artificial neural Table 3 Proposed modules (study areas) within Numerical and Data Specialist for Science Domains. Area Modules ECTS Scientific Computing 3 9 Advanced HPC Topics 3 9 AI/ML 3 9 HPDA 3 9 Project 1 9 Shared Modules 7 21 In Total 20 66 networks for science and engineering problems. Their HPC implementation uses existing software frameworks, for example, TensorFlow and PyTorch, in parallel architecture, using CPU and GPUs. High-performance data analytics It targets big data, high-dimensional data analysis, and data visualisation. Big data analysis includes file systems and databases for big data and machine learning for big data connected with its analysis using tools and programming models, such as Apache Spark and the Hadoop Map-Reduce ecosystem. With respect to high-dimensional data analysis, more explanation will be given to the geometry of high-dimensional data sets, nonlinear dimension reduction and manifold models, linear dimension reduction, principal component analysis and kernel methods, mathematical distance and dimension reduction, singular value decomposition, and principal component analysis. Finally, we introduce data handling and visualisation tools like ParaView and VisIt. Project It encourages the study group to focus either on mathematical modelling (numerical methods, optimisation, and statistical modelling) or parallel implementation and analysis of mathematical algorithms on HPC machines, especially considering real-world applications such as aerodynamics, machine learning, material science, etc. Shared modules This will help the study group integrate their knowledge with other beneficial specialisation topics, considering the area of scientific computing and, for example, covering advanced topics in large-scale computer architecture (from “System Architect”), operating systems, storage, Virtualisation & Containerisation (from “System Development and Support”) and performance analysis tools for heterogeneous HPC machines (from “Performance Analyst and Advisor”). Essentially, Journal of Parallel and Distributed Computing 201 (2025) 105081 6 P. Bouvry, M. Brorsson, R. Canal et al. Fig. 5. Proposed modules (study areas) within Performance Analyst and Advisor. Table 4 Proposed modules (study areas) within Performance Analyst and Advisor. Area Modules ECTS Performance Analysis 2 6 Performance Modelling 2 6 Performance Optimisation 2 6 Project 1 9 Shared Modules 8 24 In Total 15 51 it will give a further better understanding of the detailed study of computational science. 4.2.2. Performance analyst and advisor This specialisation targets graduate experts in performance analysis and tuning of complex application codes for modern computer architectures and supercomputers. It follows a structured performance engineering approach consisting of performance analysis based on measurements, performance modelling to estimate performance limits and bottlenecks on a certain platform and performance optimisation using hand-crafted or automatically generated architecture-specific implementations. Fig. 5and Table 4show an overview of the areas targeted and the expected ECTS within this specialisation. A detailed description of each area follows. Performance analysis The content includes criteria for selecting the right tool for specific performance analysis tasks and introduces various performance tools such as profilers, tracers, debugging tools, and monitoring tools. Students will learn to profile parallel applications to identify performance bottlenecks, perform real-time debugging and tracing of multithreaded and distributed applications, and monitor system performance and resource utilization. The modules provide an indepth exploration of profiling tools like Nvidia Insight, Intel VTune, and likwid, and debugging tools such as TotalView and DDT. Performance monitoring and tracing tools including Intel® Trace Analyzer, Scalasca, Score-P, Vampir, TAU, and Paraver, as well as monitoring tools like Ganglia, Nagios, and Prometheus are also covered. Students will create custom monitoring scripts and dashboards for HPC systems and develop optimisation strategies based on insights from performance tools, focusing on code restructuring, load balancing, and data management. Performance-driven decision-making for hardware upgrades or configurations is also addressed. Practical exercises involve hands-on experience with profiling tools to analyse compute and memory performance, using performance analysis tools to determine the performance of parallel applications, and detecting overhead and bottlenecks in sample parallel applications. Performance modelling It covers the principles of performance models, comparing techniques for validating models against existing implementations, and calibrating models to ensure accuracy and reliability. It also includes performance modelling for software applications such as web applications, databases, and parallel programs, and involves analysing workload patterns, response time, and throughput to identify performance bottlenecks using models. Students will apply performance modelling at the node and system levels, including for distributed systems, cloud services, and HPC clusters. It also covers analysing system-level performance metrics and interactions, as well as scaling and load balancing in distributed systems. Performance models will be used to identify optimisation strategies, and students will apply model-driven optimisation techniques to improve system performance and perform costbenefit analysis of optimisation decisions. Practical exercises include creating basic performance models, writing case studies of model validation, and predicting bottlenecks in parallel application samples. Performance optimisation For optimisation of a provided implementation on a specific platform, one first needs to be able to compare the architectures of modern supercomputers and understand single-core architecture and optimisation strategies. Memory hierarchy and data access optimisations, efficient shared memory parallelization, and an evaluation of different parallelization approaches for multi-core processors, including GPUs, are necessary. It also covers efficient distributed memory parallelization, advanced performance models, and the comparison of serial and parallel performance modelling. Additionally, students will evaluate energy-efficient implementation and execution of parallel programs. Practical exercises involve implementing basic numerical methods with high hardware efficiency on parallel computers and gaining insight into innovative programming techniques and alternative supercomputer architectures. This also includes advanced techniques like domain-specific languages and code generation. Project The prerequisites of this (group) project are basic performance analysis and knowledge of performance models and computer architecture fundamentals. It usually involves profiling and optimizing codes on HPC systems or industry-related projects in the area of performance engineering. Shared modules Performance engineering also requires insight into computer architecture and computational methods. Thus, this specialisation shares several optional subjects with the specialisations “System Architect” (Heterogeneous Systems and Accelerators, HPC System Design, Design Tools and Simulators, High-Level Digital Design), “System Development and Support” (Advanced Operating Systems, Compiler Design, Parallel Programming Models), and “Numerical and Data Specialist for Science Domains” (Domain Specific Programming). 4.2.3. System development and support This specialisation targets graduate experts in the development of software for supercomputers, including the programming model, compiler, operating system, performance analysis tools, and middleware technologies. More specifically, we selected the following skills: the design and implementation of programming models, compilers, compiler optimisations, toolchain, software stacks (for CPU, GPU, FPGA, CRGA), performance analysis tools, operating systems, kernel development, parallel file systems, high-speed networking, synchronization, container technologies, virtualisation technologies, integration of HPC and cloud, Journal of Parallel and Distributed Computing 201 (2025) 105081 7 P. Bouvry, M. Brorsson, R. Canal et al. Fig. 6. Proposed modules (study areas) within System Development and Support. Table 5 Proposed modules (study areas) within System Development and Support. Area Modules ECTS Operating Systems and Virtualisation 2 6 Storage and System Administration 2 6 Compilers and Parallel Programming Models 2 6 Project 1 9 Shared Modules 10 30 In Total 17 57 server administration, infrastructure setup and management, and security. Fig. 6and Table 5show an overview of the areas targeted and the expected ECTS within this specialisation. A detailed description of each area follows. Operating systems and virtualisation This area includes Advanced Operating Systems and Virtualisation and Containerization Support. They provide the students with the required design and implementation alternatives around scheduling policies, memory management, parallel I/O, device drivers, and support for virtualised environments, in order for them to design new approaches for supporting these and new features on upcoming systems. The practical side includes the development of OS drivers, and evaluating the support for virtualised environments. Storage and system administration This area covers the aspects of System Administration on HPC Environments, including protection and security aspects, application schedulers and resource managers (currently Slurm...), installation of applications, development tools, development of plans for system maintenance, user management and physical integration of the components of a supercomputer. Also, the Storage, Parallel File Systems and Databases include the design, implementation and management of parallel file systems (currently GPFS, Lustre, Gluster, etc.), the use of advanced storage devices, and parallel highperformance databases. Practical includes installation of the OS itself, and configuration of services taking into account the protection and security requirements, as well as the design and implementation of parts of a parallel filesystem. Compilers and parallel programming models This area covers the design and implementation of Parallel Programming Models and Compilers for HPC environments. It includes the design of parallel languages and models for programming HPC systems, design of high-level abstractions for expressing parallelism (currently MPI, OpenMP, Kokkos, Raja, etc.), task-based parallel programming, interoperability between models, compiler design for HPC and heterogeneous systems, and implementation of compiler optimisations. Practical covers the implementation of compiler passes for optimisations, code generation, and language and programming model extensions to improve expressiveness or performance portability. Project In this module, students work in groups with the goal of providing a solution to a particular problem related to the specialisation areas. Shared modules Additionally, this specialisation shares several optional subjects with the specialisations “System Architect” and “Performance Analyst and Advisor”. From “System Architect” the modules on “Design Tools and Simulators”, “Processor Design”, and “Multiprocessor Design”. From “Performance Analyst and Advisor”, the modules on “Tool-based Performance Engineering” and “Model-based Performance Engineering”. 4.2.4. System architect This specialisation targets graduate experts in the design and development of supercomputers. Focus on processor, multiprocessor, supercomputer architectures, memory and I/O systems, networking, circuit design, verification and test, low-power techniques, and fabrication. The next subsections describe the learning outcomes of each identified study area. Fig. 7and Table 6show an overview of the areas targeted and the expected ECTS within this specialisation. A detailed description of each area follows. System and hardware technologies This study area includes contents related to HPC system design, hardware technologies and sustainable nanoelectronic design. In detail, HPC system design groups the content around HPC facilities (e.g. power delivery, cooling solutions, use of resources), and HPC system components (e.g. CPUs, accelerators, memories, networking, racks, blades). In the area of hardware technologies, we included the presentation of the underlying silicon and photonics technologies (MOS, CMOS, memories, photonics), Moore’s law and Dennard scaling, power wall and dark silicon, system reliability (soft errors, hard errors, solutions) and Security (Root of trust, encryption primitives). Finally, in the Sustainable nanoelectronic design, we include the principles of sustainability in integrated circuits, sources of power dissipation, low-power techniques, lifetime analysis and optimisation, CO2 footprint; and the societal and economic impact. CPUs and heterogeneous systems This study area includes content related to processor design, multiprocessor design, high-level digital design, and heterogeneous systems and accelerators. Specifically, processor design includes CPU pipelining, memory hierarchy, caches, branch prediction, reservation stations, reorder buffer, load-store queue and hardware performance counters. Multiprocessor Design focuses on the logic needed for data coherency and consistency, the on-chip network and the memory hierarchy. Then, high-level digital design is devoted to the principles of digital systems based on programmable and configurable components, the system breakdown into hardware and software components and the design of processing subsystems (e.g., video, vector and matrix operations, sorting algorithms). Finally, heterogeneous systems and accelerators include single chip, multiple chip (chiplet) and Journal of Parallel and Distributed Computing 201 (2025) 105081 8 P. Bouvry, M. Brorsson, R. Canal et al. Fig. 7. Proposed modules (study areas) within System Architect. Table 6 Proposed modules (study areas) within System Architect. Area Modules ECTS System and Hardware Technologies 3 9 CPUs and Heterogeneous Systems 4 12 Chip Design and Test 3 9 Project 1 9 Shared Modules 5 15 In Total 16 54 multiple board architectures and accelerators (e.g. Vector unit, GPU, FPGA, TPUs, and stencils). Chip design and test This study area groups the content related to Physical Design and Test and the related tools. In detail, System-on-Chip Physical Design includes the timing and power constraints to a complex integrated circuit, the physical implementation of a complex integrated circuit, low power design techniques, design, analysis and evaluation of electronic systems in applications such as automation, energy distribution and generation, consumer electronics, biomedicine, etc. In the module dedicated to the testing of digital designs, we include fault models, fault coverage, observability, automatic test pattern generators, scan-based tests and self-tests. Finally, in the area of Design Tools and Simulators, students will be presented and use HDLs (e.g., SystemC, SystemVerilog) to describe systems, apply the VLSI design flow to hardware design and use physical design tools (partitioning, chip planning, placement, global routing, detailed routing). Project In this module, we expect students to select a certain topic related to ongoing research projects, running compute-time projects on HPC systems, industry-related projects in the area of system architecture, or performing hands-on system development, modelling or optimisation. Shared modules Additionally, this specialisation shares several optional subjects with the specialisations “System Development and Support” and “Performance Analyst and Advisor”. From “System Development and Support”, the modules on Advanced Operating Systems, Storage, Parallel File Systems and Databases and Compiler Design are highly interesting for this specialisation together with the Tool-Based Performance Analysis in the “Performance Analyst and Advisor” specialisation. 4.3. Transversal skills The EUMaster4HPC curriculum is designed to include in addition to fundamental HPC skills also transversal skills. These transversal skills are essential for students to succeed in their careers and to prepare them effectively for the evolving job market. We adopt the following definition of transversal skills by the UNESCO International Bureau of Education: “Skills that are typically considered as not specifically related to a particular job, task, academic discipline or area of knowledge and that can be used in a wide variety of situations and work settings (for example, organisational skills)” [17]. A subset of these skills involves soft skills (sometimes also called professional skills): “Term used to indicate a set of intangible personal qualities, traits, attributes, habits and attitudes that can be used in many different types of jobs”. Nowadays, many top technology companies test applicants on their soft skills during the interview process. Taking into account this attitude, a set of advanced skills is proposed within the curriculum to train students to address issues such as diversity, inclusion, gender, and unconscious bias. Based on the method described in Section 3, six groups of skills have been identified, which are listed below. For each of these groups, modules have been defined in terms of content topics: •Research Techniques: scientific methods, critical thinking, curiosity, artificial intelligence in higher education. •Scientific Writing: writing papers, publishing scientific results, scientific publications databases, technical presentations, and proposal writing. •Work Environment: ethics and integrity, team management, human resources, relational skills (communications), priority management, gender, diversity. •Entrepreneurship: business organization and management, business plan development, market viability, financial management, (public) procurement procedures, intellectual property, and identifying customer needs. •Legal Aspects: licensing, open data, intellectual property, data protection and privacy (GDPR), EU regulations in information technology. •Other: attention to detail and aesthetics, curiosity, willingness to raise questions and to rely on the expertise of other people, ambition to acquire more skills and willingness to continuously learn new things. 4.4. Balancing modules Most of the time, high-performance computing is misinterpreted as related to pure computer science. However, this is not true; from the early stages to the present day, computers have been used for doing arithmetic computations using large volumes of data sets. For example, solving a large number of systems of equations, AI computations using large volumes of data sets, material science and biomedical simulationbased specific mathematical modelling and computer graphics. All of those examples require broad knowledge of physics, mathematics, engineering, and computer science skills. Therefore, to become a highly qualified person in high-performance computing, one must have more than intermediate knowledge in physics, mathematics, and engineering. Therefore, it is essential that we encourage students with backgrounds in physics, engineering, and mathematics to study master’s in HPC. However, those who have bachelor’s degrees in engineering, physics, and mathematics may lack knowledge of pure computer science. Journal of Parallel and Distributed Computing 201 (2025) 105081 9