Speaker:

Dr. Johannes Langguth

Simula, Norway

Johannes Langguth is a research scientist at Simula research laboratory, Oslo, Norway. He received his PhD in computer science from the University of Bergen in 2011, followed by a postdoctoral appointment at ENS Lyon, France. His research interests include computer architecture, parallel algorithms, computational social science, and high-performance scientific computing on multi-core CPUs and GPUs.

Detailed organ-scale simulations of calcium handling and electrical signal transmission in the human heart require stochastic simulation of a large number of ion channels in each cell, which consumes immense processing power for the simulation of a single heartbeat, thereby creating the need for large scale parallel implementations. We present codes for solving such cardiac models on structured and unstructured meshes, and discuss the challenges involved in modernizing these codes to run on heterogeneous supercomputers. We focus on the interaction between OpenMP, MPI, and CUDA in such computations, as well as optimizations to communication and vector processing, and illustrate practical experiences with these applications on different supercomputers.

Speaker:

Prof. Eric Aubanel

University of New Brunswick, Canada

Dr. Eric Aubanel is a Professor in the Faculty of Computer Science at The University of New Brunswick, Canada. He received his Bachelors of Science from Trent University and his PhD from Queen's University. Eric's current research includes the development of tools and algorithms to support scientific computing on heterogeneous distributed resources, including manycore accelerators. He is also beginning a study of the cognitive aspects of parallel programming. In 2016 he published a graduate textbook on parallel computing, Elements of Parallel Computing, with CRC Press. He is a member the IBM CASA (Centre for Advanced Studies - Atlantic), where he has participated in projects on performance optimization of managed runtimes.

Realistic simulation of time-dependent processes requires long time evolution using very small time steps in order to resolve phenomena at different time scales. Traditional spatial parallelism offers some performance gain, which however may be insufficient. The parareal algorithm parallelizes partial differential equations by time decomposition, and has been employed for a range of physical problems. It faces particular challenges in the case of highly nonlinear dynamics, where the parareal corrections can cause significant instability. This talk describes strategies that have been used to deal with this instability, including grid coarsening techniques, filtering of grid-dependent features, and multistage convergence techniques, in the context of turbulent flow in computational fluid dynamics.

Speaker:

Prof. Ioana Banicescu

Mississippi State University, USA

Dr. Ioana Banicescu is a professor in the Department of Computer Science and Engineering at Mississippi State University (MSU). Between 2010 and 2017, she was also a Director of the Center for Cloud and Autonomic Computing at MSU, and a Co-Director of the National Science Foundation Center for Cloud and Autonomic Computing. She received the Diploma in Engineering (Electronics and Telecommunications) from Polytechnic University - Bucharest, and the MS and the PhD degrees in Computer Science from New York University - Polytechnic Institute. Professor Banicescu's research focus is on performance optimization for problems in computational science, autonomic computing and graph analytics. Her research interests include parallel algorithms, scientific computing, scheduling, and performance modeling, analysis and prediction. Professor Banicescu is the recipient of a number of awards for research and scholarship from the National Science Foundation. She serves on numerous scientific panels for advanced research grants in the US and Europe, on journal editorial boards, and on the steering and program committees of a number of international conferences, symposia and workshops. She served on the Executive Board and Advisory Board of the IEEE Technical Committee on Parallel Processing (TCPP).

Computational problems in science and engineering are continuously increasing in size and complexity, and their solutions can often be made tractable only by using parallel and distributed computing environments with state-of-the-art techniques and tools. Such computational environments are prone to irregular behavior due to unpredictable variations in problem, algorithm and system characteristics. In situations where application execution times are stochastic in nature and the availability of resources is uncertain, a robustness study of resource allocations and application scheduling is required to guarantee a desired level of performance. This talk will reveal highlights of the past and ongoing work on such a robustness study, including models, heuristics, and a framework for robust allocations and scheduling of scientific applications in heterogeneous environments. Moreover, robustness prediction and evaluation will be described for scheduling arbitrarily divisible workloads for computational intensive and communication intensive scientific applications, thus underscoring the power of analytical prediction of the robustness of resource allocations and application scheduling in various computational environments. The significance of the robustness study using stochastic models in providing a cost-effective and low overhead analysis of robust allocation and scheduling will also be discussed.