Reproducible science is contingent on the ability to execute a workflow once, and easily share software and workflow dependencies to run again. This presents a significant problem in the case of disparate environments, as there is large variance in executors afforded by such environments. Snakemake, enables the scientific user to represent an entire workflow in a transparent and readable way, while providing seamless execution on single computers, cluster environments and the cloud [1].
Speaker Description:
Vanessa Sochat is a research software engineer in the Stanford Research Computing Center.
HPC support teams are often tasked with installing scientific software for their user community and the complexity of managing a large software stack gets very challenging. Software installation brings forth many challenges that requires a team of domain expertise and countless hours troubleshooting to build an optimal software state that is tuned to the architecture. In the past decade, two software build tools (Easybuild, Spack) have emerged that are widely accepted in HPC community to accelerate building a complete software stack for HPC systems.
Speaker Description:
Shahzeb Siddiqui started out his career in High Performance Computing (HPC) in 2012 at King Abdullah University of Science and Technology (KAUST) while pursuing his Masters. His focus in HPC includes Parallel Programming, Performance Tuning, Containers (Singularity, Docker), Linux system administration, Scientific Software Installation and testing, Scheduler Optimization, and Job Metrics. Shahzeb has held multiple roles in his HPC career in the following companies: Dassault-Systemes, Pfizer, Penn State, and IBM. Prior to 2012, he was a software engineer holding multiple roles at Global Science & Technology, Northrop Grumman, and Penn State.
Shahzeb is the creator of open-source project buildtest a Software Stack Testing Framework designed to automate testing for HPC systems. Shahzeb is an experienced Developer, Dev-Ops, Cloud Engineer, and System Administrator. He holds a M.S in Computer Science from KAUST and B.S in Computer Engineer from Penn State University.
Software practices in industry are subjected to a ruthless Darwinian selection. Current best practices represent the least expensive and most effective way to safely develop reliable software.
Many large scientific software systems, even those in operational use, do not use these proven tools. Efforts to improve process are often bogged down by technical difficulties, management neglect, or programmer resistance.
Speaker Description:
I am the author of netCDF-4, currently also working on the PIO library for HPC I/O. I have managed many software teams in science and industry.
Producing scientific software is a challenge. The high-performance modeling and simulation community, in particular, is dealing with the confluence of disruptive changes in computing architectures and new opportunities (and demands) for greatly improved simulation capabilities, especially through coupling physics and scales. At the same time, computational science and engineering (CSE), as well as other areas of science, are experiencing increasing focus on scientific reproducibility and software quality.
Speaker Description:
David Bernholdt is a Distinguished R&D Staff Member and Group Leader at Oak Ridge National Laboratory (ORNL). He has leadership roles in multiple projects in the DOE Exascale Computing Project (ECP) and the Scientific Discovery through Advanced Computing (SciDAC) program. He also leads the Programming Environment and Tools area of the Oak Ridge Leadership Computing Facility (OLCF). His research interests center on making it easier and more productive to create and use computational science and engineering software on the largest high-performance computer systems.
Machine learning, and specifically artificial neural networks, are powerful and dangerous tools that have become too easy to implement, and too difficult to understand. I am proposing running a tutorial to build up a feedforward artificial neural network from scratch. The tutorial will start with a brief history on the beginnings of artificial neural networks and how they originated out of trying to understand the brain. We will briefly discuss the ‘AI winter’, and the breakthrough algorithm backpropagation which brought us back into the ‘AI summer’.
Speaker Description:
I am an Associate Research Scientist at RadiaSoft, LLC working on machine learning for various applications, including prostate cancer treatment plans and particle accelerators. I recently completed my PhD In Computational Bioscience with thesis work on the intersection of machine learning and neuroscience.
With the increasing complexity and diversity of the software stack and system architecture of high performance computing (HPC) systems, the traditional HPC community is facing a huge productivity challenge in software building, integration and deployment for multiple exascale computing systems that will be deployed in year 2020 and after. Recently, this challenge has been addressed by new software build management tools such as Spack that enable seamless software building and integration. Container based solutions provide a versatile way to
Speaker Description:
Sameer Shende serves as the Director of the Performance Research Laboratory at the University of Oregon. His research interests include performance evaluation tools, runtime systems, instrumentation, measurement, and analysis tools, and optimizing compilers.
Deep learning is rapidly and fundamentally transforming the way science and industry use data to solve problems. Deep neural network models have been shown to be powerful tools for extracting insights from data across a large number of domains. As these models grow in complexity to solve increasingly challenging problems with larger and larger datasets, the need for scalable methods and software to train them grows accordingly.
Speaker Description:
Steve Farrell is a Machine Learning Engineer at the NERSC supercomputing center at Lawrence Berkeley National Laboratory. In this role he supports the ML needs of 7000 users across a wide range of scientific domains, and the ML software stack. He collaborates with science teams to perform applied ML research.
Modern science is getting bigger in every way: the scope of the problems, the size of the datasets, and the required computing power. Massive undertakings such as the Large Hadron Collider and the Large Synoptic Survey Telescope as well as major looming challenges such as global climate change are pushing science past its limits. Deep Learning, a methodology that has seen tremendous success in industry applications, is poised to transform scientific workflows.
Speaker Description:
Steve Farrell is a Machine Learning Engineer at the NERSC supercomputing center at Lawrence Berkeley National Laboratory. In this role he supports the ML needs of 7000 users across a wide range of scientific domains, and the ML software stack. He collaborates with science teams to perform applied ML research.
Model uncertainty estimation using the ensemble approach poses a large computation requirement. This is because ensembles are usually generated form multi-model and multi-simulation with slightly perturbed initialization.
Speaker Description:
Weiming Hu is a Ph.D. student of Prof. Guido Cervone at Penn State University in the Dept. of Geography focusing on computational algorithms, and numerical weather prediction; Guido Cervone is the Associate Professor at Dept. of Geography and the Associate Director of Institue for CyberScience at Penn State University.
He can be reached at weiming@psu.edu if you have any questions.
Deep Learning model complexity and training data volume continue to grow rapidly. Training with a single-GPU, or even a single-node of GPUs, is often too slow for the iterative nature of model development and optimization. In this talk, we will discuss several aspects of building scalable GPU-based clusters for improving training time including scaling training to multiple-nodes, optimizing inter-node collective operations, and optimizing the data-cache hierarchy.
Speaker Description:
Craig Tierney is a Senior Solution Architect at NVIDIA supporting high performance computing (HPC) and deep learning (DL). His focus includes the architecture of GPU based systems to maximize HPC and DL performance and scalability. Prior to joining NVIDIA, Craig spent over 15 years providing high performance computing architecture and computational science support to NOAA and several other government and educational organizations including DOE, DOD, NASA and Stanford University. Craig holds a Ph.D. in Aerospace Engineering Sciences from the University of Colorado at Boulder.
