Full Program

DescriptionAdvances in computer science have brought us to the gates of the exascale era. In order to master this massive computational power, both industry and academia are investing their efforts in developing and improving new software tools and techniques. Some of these have been quickly adopted and widely used by the community, like big data and machine learning technologies. Many scientific domains benefit from the aforementioned software, which helps expanding knowledge in their study areas. For example, using big data technologies gives scientists the opportunity to tackle complex scientific questions that otherwise would be impossible to address. Furthermore, the combination of big data and machine learning has opened a whole new field of scientific possibilities. This minisymposium explores how different scientific domains (life sciences, physics) apply big data technologies and deep learning to bring their research one step forward. We focus on those cases where it would be impossible to address scientific questions without using big data, machine learning or a combination of both. We also emphasize the need of foreseeing the technical requirements behind the experiments and how the computational ecosystem needs to be in place well in advance to ensure the success of the scientific study.

DescriptionSerguei Maliar's paper applies deep learning to macroeconomic models. The goal of his paper is to demonstrate that deep learning techniques can be used to analyze rather complex economic models in a simple and general manner. They show how to cast the typical dynamic economic model into a form that is suitable for deep learning analysis, and how to design a version of deep learning algorithm that can construct a numerical solution to the model. Xavier Ragot will solve for optimal Ramsey policies in heterogeneous-agent models with aggregate shocks. They provide a new simple theory based on projection on the space of idiosyncratic histories, to present a finite-dimensional state-space representation. Gregor Boehl will present a solution method and a Bayesian smoother to filter and estimate rational expectations models with occasionally binding constraints quickly and accurately. Ralph Luetticke's paper describes a method for solving heterogeneous agent models with aggregate risk and many idiosyncratic states formulated in discrete time. It extends the method proposed by Reiter (2009) and complements recent work by Ahn et al. (2017) on how to solve such models in continuous time.

DescriptionThe aim of this minisymposium is to discuss the latest research at the intersection of parallel computing and high-dimensional approximation. High-dimensional approximation drives e.g. uncertainty quantification and machine learning as well as big data and simulations of complex physics models. It is well-known that approximation of functions of growing dimension has the challenge of the curse of dimensionality. Over the last decades, many powerful mathematical tools have been developed to do weaken or overcome this. These include, but are not limited to sparse approximation, (quasi) Monte Carlo, multi-level (multigrid) / multi-fidelity techniques, (sparse) tensor product and low-rank approximations, hierarchical matrices, compressed sensing and meshfree methods. There is a growing interest to solve high-dimensional approximation problems at large scale. While many of the discussed methods have good or even optimal approximation properties and complexities for larger dimensions, some have been primarily developed for sequential execution. However, to solve large scale approximation problems, it becomes necessary to develop fast, scalable and parallel numerical methods. This minisymposium includes contributions in high-dimensional approximation ranging from initial studies for the use of parallel techniques up to full scale parallel methods that run on large HPC clusters. We focus both on algorithmic-oriented and application-centered research.

DescriptionThe Pangeo community is an international community committed to fostering collaboration around the open source scientific Python ecosystem for ocean / atmosphere / land / climate science, supporting the development with domain-specific geoscience packages, and improving the scalability of these tools to handle petabyte-scale datasets on HPC and cloud platforms. The Pangeo ecosystem of Python packages include Dask, Xarray, Iris, and Jupyter, and developing interoperability around these packages has led to significant advancements in the last year in the areas of Xarray- and Iris-compatible data cataloging and discovery software, Jupyter Notebook dashboarding, cloud-friendly data formats for analysis, hybrid HPC-cloud technologies and environments, just to name a few. In this session, we cover some of the recent advancements made by the Pangeo community.

DescriptionThe emergence of diverse and complex massively parallel heterogeneous hardware solutions have a large impact on the programming models traditionally used in existing large and complex weather and climate models that may well be decades old and not easily adaptable to new and possibly multiple programming models. Porting existing large community codes to multiple architectures is a daunting task and may lead to codes with multiple more complex, and more difficult to maintain variations depending on hardware choices. In order to increase the productivity and maintainability, while retaining a high degree of performance in this emerging landscape, a complete rethink of design choices is required based on abstractions and separation of concerns. This approach implies large rewriting efforts of existing codes with a considerable emphasis on low-level infrastructure developments, and automatic code generation. As a result in the past years numerous new technologies and approaches are emerging in order to provide new programming models and abstractions for concurrency and data interpretations. This minisymposium will provide an update both on infrastructure developments, source-to-source translation tools and domain-specific languages (DSL) as a way to address the challenge of bridging the software productivity gap for weather and climate models.

DescriptionEarth science relies on numerical methods now more than ever before. Dealing with the ever-growing massive observation data, simulating earthquakes or geodynamic problems, imaging the interior of earth at the finer scale, solving geological engineering issues with complicated settings, etc., all these studies benefit from the development of high-performance computing. To embrace the big-data era and harness the power of high-performance computing (HPC), the goal of this minisymposium is to discuss the scientific achievements and the theoretical/technical improvements of numerical methods we have made with HPC, as well as the emerging new frontiers like AI in computational geosciences. The minisymposium also provide a chance for computer scientists and earth scientists to work together to address the challenges in the future of Earth sciences.

DescriptionThe aim of this minisymposium is to discuss the latest research at the intersection of parallel computing and high-dimensional approximation. High-dimensional approximation drives e.g. uncertainty quantification and machine learning as well as big data and simulations of complex physics models. It is well-known that approximation of functions of growing dimension has the challenge of the curse of dimensionality. Over the last decades, many powerful mathematical tools have been developed to do weaken or overcome this. These include, but are not limited to sparse approximation, (quasi) Monte Carlo, multi-level (multigrid) / multi-fidelity techniques, (sparse) tensor product and low-rank approximations, hierarchical matrices, compressed sensing and meshfree methods. There is a growing interest to solve high-dimensional approximation problems at large scale. While many of the discussed methods have good or even optimal approximation properties and complexities for larger dimensions, some have been primarily developed for sequential execution. However, to solve large scale approximation problems, it becomes necessary to develop fast, scalable and parallel numerical methods. This minisymposium includes contributions in high-dimensional approximation ranging from initial studies for the use of parallel techniques up to full scale parallel methods that run on large HPC clusters. We focus both on algorithmic-oriented and application-centered research.

DescriptionPython has become one of the predominant programming languages among scientists due to its simplicity and a very solid ecosystem of scientific libraries for data handling, computing and visualization. As Python keeps gaining popularity, more and more efforts aim to move it from just prototyping and workflow management, to also run production applications on large HPC systems. This minisymposium gathers people working on domain specific Python frameworks for HPC applications and systems, so as to start a conversation on the possibilities of Python as an effective language for high quality HPC applications. We wish this minisymposium to be a starting point to reason about code maintainability and scalability (code size/productivity/debuggability/profiling), deployment models on HPC systems, optimization opportunities, programing models, but also to highlight the opportunities that the Python ecosystem could offer for the next generation of parallel systems and applications, like interactive scientific applications. The talks will present different on-going efforts from several authors to scale actual scientific frameworks to large HPC systems, using different techniques, such as handling of multi-processes and threads and code generation. This will allow us to assess advantages and disadvantages of the different solutions and have a better understanding of the challenges ahead.

DescriptionRecent advances in theory and numerical methods in parallel to the availability of high-quality massive data sets and high-performance computing provide unprecedented opportunities to improve our understanding of Earth's interior and its mechanism. The goal of this session is to bring computational and Earth scientists together to form a platform to discuss the current status, challenges and future directions in computational seismology and computational Earth Sciences. Talks highlight advances through numerical (high-performance computing) simulations, their algorithms and their scientific outcomes. They help to identify opportunities and trends made possible through better computers and data, and they help to uncover challenges and problems arising from better data, computer architecture changes and more complex software. Contributions include, but are not limited to, the areas of earthquake engineering, passive and active-source seismic imaging, links to geodynamical modelling, observational data and laboratory experiments in conjunction with computational approaches such as numerical solvers, large-scale workflow, big data, optimisation strategies and machine-learning on HPC systems.

This session is dedicated to Dimitri Komatitsch.

Description Fortran is one of the most prominent languages for scientific applications in the High Performance Computing community. In some cases, the developments of these applications have been started more than 20 years ago, and they are still actively developed and used by large scientific communities. For some science domains, such as material science and astrophysics, they represent the most consuming applications in terms of computing resources in several computing centers. For this reason, it becomes crucial to extend the existing approaches to application programming, in particular with the approach of the exascale era, when a disruption will be likely necessary for application design to achieve effective exascale performance levels. Although most of the codes make use of old Fortran standard features (mainly F95), an effort of modernization of the language itself has been undertaken with newer revisions of the standard (F03/F08, F18), some of them by introduction of new software engineering techniques. In this minisymposium we will discuss, with invited presentations by prominent experts in the field, how the combination of the new Fortran standard features and software engineering techniques can help to face the exascale challenges for existing Fortran applications.

DescriptionThe most comprehensive description of plasma phenomena involves kinetic theory, with applications in different fields, such as magnetic or inertial fusion, industrial plasmas and astrophysics. Solving kinetic problems requires overcoming multiple challenges: the high dimension of phase space (6D), time and spatial scales spanning several orders of magnitude, various non-linearities and the need to solve the combined system of Vlasov/Fokker-Planck equations consistently with Maxwell equations. Reduced models have been devised tailored for a particular application: for example gyrokinetic theory for turbulent transport, or hybrid fluid/kinetic models for fast particle effects on MHD modes. Nevertheless, the resulting equations remain numerically very challenging and, with the aim of carrying out ever more realistic simulations, require running on the most powerful computers available at any time. To make efficient use of these HPC resources, and to maintain readiness for future developments, efforts are being pursued at multiple levels: Refactoring and porting legacy codes on new and emerging architectures; Improving code design for facilitating future portability and further implementation of physics models of increased complexity; Developing improved algorithms as well as innovative numerical approaches running efficiently on future HPC platforms.

DescriptionWith limitations in expansion of the von Neumann architecture processing units, several new models for running computations are gaining attention from the high energy physics community. These exotic architectures are considered potential platforms of scientific computation. Even though they are not foreseen to be used in the near future, the community needs to be aware of their potential and prepare for when new accelerators might become commercially viable and available. These new architectures require new programming paradigms and somehow limit the type of algorithms they can run. We propose to discuss the usage of non-standard computing architectures as accelerator to part of the high energy physics applications. This includes, but is not limited to, applications running on field programmable gate arrays, neuromorphic architectures, and quantum chips. With this minisymposium, we hope to create a picture of the state of the art, raise awareness of these architectures and guide future developments.

DescriptionRecent advances in theory and numerical methods in parallel to the availability of high-quality massive data sets and high-performance computing provide unprecedented opportunities to improve our understanding of Earth's interior and its mechanism. The goal of this session is to bring computational and Earth scientists together to form a platform to discuss the current status, challenges and future directions in computational seismology and computational Earth Sciences. Talks highlight advances through numerical (high-performance computing) simulations, their algorithms and their scientific outcomes. They help to identify opportunities and trends made possible through better computers and data, and they help to uncover challenges and problems arising from better data, computer architecture changes and more complex software. Contributions include, but are not limited to, the areas of earthquake engineering, passive and active-source seismic imaging, links to geodynamical modelling, observational data and laboratory experiments in conjunction with computational approaches such as numerical solvers, large-scale workflow, big data, optimisation strategies and machine-learning on HPC systems.

This session is dedicated to Dimitri Komatitsch.

DescriptionThe most comprehensive description of plasma phenomena involves kinetic theory, with applications in different fields, such as magnetic or inertial fusion, industrial plasmas and astrophysics. Solving kinetic problems requires overcoming multiple challenges: the high dimension of phase space (6D), time and spatial scales spanning several orders of magnitude, various non-linearities and the need to solve the combined system of Vlasov/Fokker-Planck equations consistently with Maxwell equations. Reduced models have been devised tailored for a particular application: for example gyrokinetic theory for turbulent transport, or hybrid fluid/kinetic models for fast particle effects on MHD modes. Nevertheless, the resulting equations remain numerically very challenging and, with the aim of carrying out ever more realistic simulations, require running on the most powerful computers available at any time. To make efficient use of these HPC resources, and to maintain readiness for future developments, efforts are being pursued at multiple levels: Refactoring and porting legacy codes on new and emerging architectures; Improving code design for facilitating future portability and further implementation of physics models of increased complexity; Developing improved algorithms as well as innovative numerical approaches running efficiently on future HPC platforms.

DescriptionNumerical weather prediction and climate models are complex scientific applications that need to run on large parallel computer systems. The rapid change of computing architectures and the increasing diversity of architectures and programming models required to run these applications presents a significant challenge for the modelling community to retain single source codes that run efficiently on multiple architectures, now and in the future. In order to address the performance portability problem, numerous solutions have emerged in recent years: i.e. source-to-source translators, domain specific languages (DSLs) and libraries that abstract details of the efficient implementation of the physical equations. However, each of these solutions applies to a particular domain, and supports certain types of horizontal grids, numerical methods or computational pattern. Reuse of abstractions and tools, such as optimizers, for the weather and climate domain, will be key to enable the sustainability and maintainability of the ecosystem of tools and libraries for weather and climate prediction applications. The standardization of interfaces will allow higher level abstractions to be built which will, in turn, increase scientific productivity by improving the ability to develop models. We will present and discuss recent efforts that have begun on the interoperability and standardization of these abstractions.

DescriptionAs the most important tools for ocean and climate process understanding, forecasting, and projection, the ocean and climate models have achieved great progress in last decade. The main stream of ocean and climate model development focuses on higher resolution, more accurate parameterizations of unresolved physical processes, and including more biogeochemical processes. At the same time, exascale high-performance computing (HPC) has gradually matured. How to use supercomputers more efficiently, such as multi core processor, hyperthread, large cache, high bandwidth process communication structure and high speed I/O functionality, becomes a critical condition for model development now. Meanwhile, the development of information technologies has led to a new era in computation, affecting almost all fields in science and engineering. Recently, increasing with large amounts of simulation and observation data, machine learning, especially deep learning shows its potential on ocean and climate simulations, such as subgrid parameterization schemes, ENSO prediction, typhoon track forecasting, etc. The convergence of simulation and deep learning brings challenges and opportunities. This minisymposium encompasses the interdisciplinary research on ocean, climate simulation/forecasting/prediction/projection and deep learning - from algorithms and new frameworks through to exciting results, outlooks and perspectives.

DescriptionThis workshop presents numerical models based on modern computational techniques for the simulation of various tissues and organs, in relation to biomedical applications. The purpose is to get a better description of still poorly understood physiological processes and/or to provide the clinical community with new tools for treatment planning and diagnosis.

DescriptionResearch and industry invest considerable efforts into using high performance computing to advance machine learning. In this mini-symposium, we propose to explore the confluence of HPC and Machine learning in the opposite direction, and ask: How can machine learning advance high performance computing? The expanding landscape of complex and heterogeneous hardware, programming paradigms and tools, means more expertise is needed to develop optimal programs for supercomputing systems and make them portable. Simultaneously though, advances in machine learning and natural language processing, as well as the large amounts of code made available by open-source repositories present opportunities in tackling problems previously unsolved using machine learning. Four talks will present methods that use machine learning in current HPC systems or applications for: performance modelling and optimization; autotuning; automatic code analysis; and optimized compilation. By showcasing these research topics and bringing together researchers from both the HPC and the ML communities, we intend to raise awareness, in the HPC community, of solutions machine learning offer to rising HPC problems as well as to create a platform for exchange among actors at the interface between HPC and ML.