Pacific Northwest National Laboratory

End-to-End Agile Hardware Generation for Sparse Tensor Workloads
Souradip Ghosh, Carnegie Mellon University
Practicum Year: 2023
Practicum Supervisor: Antonino Tumeo, Chief Scientist, HPC Team B, Pacific Northwest National Laboratory
In the age of specialized accelerators, *agile* hardware generation is critical to speed up the development cycle from application to hardware design to silicon. SODA-OPT is a framework that provides end-to-end agile hardware generation from high-level source code (e.g. Python, TensorFlow, etc.) via interchangeable HLS techniques and completely automated compilation (optimization and lowering). While SODA-OPT has been successful in fully automating hardware synthesis for regular ML models (in TensorFlow and PyTorch), it has struggled to provide the same automation for irregular workloads: namely sparse tensor algebra and graph analytics, more broadly. One of the major bottlenecks in automating hardware generation is the lacking compilation support in an end-to-end HLS framework. This project explores how sparse tensor algebra, graph analytics, and irregular workloads should be expressed in the upper end of the stack (i.e. at the language level and in the compiler's intermediate representation/IR) so they can be optimized and lowered for and during HLS.
Snow drought characteristics in E3SM
Marianne Cowherd, University of California, Berkeley
Practicum Year: 2022
Practicum Supervisor: Ruby Leung, Battelle Fellow, Earth Science, Earth Systems Analysis & Modeling, Pacific Northwest National Laboratory
I used model outputs from E3SM to study how the model represents the past frequency and severity of snow droughts and how those rates will change in the future. In particular, I worked on classifying droughts as either warm or dry, depending on their cause within the model.
Protonation of Serine in Gas and Condensed and Microsolvated States in Aqueous Solution
Steven Wilson, Arizona State University
Practicum Year: 2022
Practicum Supervisor: Samantha Jo Johnson, Computational Scientist, Physical and Computational Sciences Directorate, Pacific Northwest National Laboratory
In order to reduce the time- -intensive nature of these calculations we propose the use of a ML algorithm to connect static calculation to the dynamic MD simulations. In this way, we hope to connect to wavefunction- -based descriptors like magnetic moments, charge density distribution, and non- -covalent interactions obtained via static DFT to the dynamic behavior of the system obtained via MD. The ability to elucidate both the electronic structure and dynamic characteristics of an analyte from a one stop turnkey operation will help quickly identify analyte characteristics and broaden our understanding of the relationship between MS patterns and electronic structures.
Emulation of warm rain processes using deep neural networks
Jamin Rader, Colorado State University
Practicum Year: 2021
Practicum Supervisor: Po-Lun Ma, Global Atmos Modeling Lead / Earth Scientist, PNNL Earth System Modeling, Pacific Northwest National Laboratory
Autoconversion, the collision of cloud droplets to form drizzle droplets, is simulated in global climate models using simple parameterizations. These simple parameterizations are applied generally to the whole globe, but the relationships that dictate autoconversion differ between various cloud/climate regimes (e.g. subtropical marine stratocumulus versus tropical deep convection). This project seeks to identify ways to replace these simple parameterizations with deep neural networks (DNNs) that can use knowledge of the atmospheric state to provide accurate estimates of the autoconversion rate. These DNN emulators will be used to improve simulations of precipitation in future versions of the DOE's Energy Exascale Earth System Model (E3SM).
MLIR Dialect for Generic Dataflow
Amalee Wilson, Stanford University
Practicum Year: 2021
Practicum Supervisor: Roberto, Gioiosa, HPC group (CS and Mathematics division), Pacific Northwest National Laboratory
Dataflow architectures have been shown to be more energy efficient than conventional CPUs and GPUs, and utilizing these architectures for accelerating AI applications, particularly sparse AI models, is an active area of research. While extremely promising for energy-efficient AI acceleration, these architectures are much more difficult to program. Exacerbating this issue is the richness of AI programming frameworks and availability of different architectures: without a common representation of the program, each programming framework must be mapped and tediously remapped to each new architecture. This work creates a generic dataflow intermediate representation using the popular MLIR framework, which was initially created to support the progressive lowering of dataflow graphs to optimized target-specific code. Many programs written in different AI frameworks (e.g. PyTorch through NPComp and Tensorflow) can already be compiled to or lowered to MLIR. The goal of the generic dataflow dialect is to support lowering from these frameworks, performing (dataflow-specific) architecture-agnostic optimizations, and lowering to optimized code for each target.
Ocean Dynamics in the GeoMIP G1 Suite of GCM Simulations
Hansi Singh, University of Washington
Practicum Year: 2014
Practicum Supervisor: Phillip Rasch, Chief Scientist for Climate Science, Atmospheric Sciences and Global Change, Pacific Northwest National Laboratory
Analyzed model output from the GeoMIP suite of climate models regarding the effect of geoengineering via solar radiation management on the ocean. Worked on attribution of the ocean response to the changes in the surface wind stress forcing.
Genetic algorithm-based inference of protein-protein interactions
Jeffrey Kilpatrick, Rice University
Practicum Year: 2010
Practicum Supervisor: Haluk Resat, Senior Research Scientist, Biological Sciences, Pacific Northwest National Laboratory
The complexity and robustness of biology is engineered in a similar fashion to many man-made large-scale systems: with a limited number of parts interacting in multiple contexts to achieve various goals. Proteins, which are themselves comprised of domains that often occur in many molecules and species, form the basis of this machinery. We extended a framework to infer protein-protein interactions by way of guessing the strength of domain-domain interactions using a genetic algorithm. We found that the procedure produces robust predictions, even when training with data from distantly related species.
Implementation of coupled-cluster linear response theory for molecular properties within NWChem
Jeff Hammond, University of Chicago
Practicum Year: 2006
Practicum Supervisor: Dr. Bert deJong, Senior Research Scientist, EMSL: MSCF Visualization & User Services, Pacific Northwest National Laboratory
The goal of this project was to add molecular property functionality to the high-accuracy methods within NWChem. Karol Kowalski and I implemented linear-response theory for CCSD and CCSDT dipole moments and dipole polarizabilities. This code allows the study of molecules previously inaccessible with these levels of theory. We are currently extending these developments in a number of directions, including QM/MM calculations of solvated molecules in collaboration with Marat Valiev and other kinds properties with Bert deJong.
Integration of models TNF and EGF signaling pathways in HMEC cells
Bree Aldridge, Massachusetts Institute of Technology
Practicum Year: 2004
Practicum Supervisor: Steven Wiley, Director, Biomolecular Systems, Systems Biology, Biomolecular Systems, Pacific Northwest National Laboratory
The purpose of the project was to lay the computational foundation for merging large scale signaling pathway models. The growing number of ODE-based models of signaling pathways being developed at PNNL and in academia creates the need to explore how to combine and compare large models. Over the course of the practicum, we re-formatted three large (500-1000 ODEs) models to be SMBL-exportable and graphically supported. To prepare exploring how the three models might be merged, a series of quantitative Western blots were performed. We established that before merging the models together, we need a mathematical representation of extracellular crosstalk. The Western blots also helped us clarify the downstream crosstalk between TNF and EGF receptor signaling. Collaborative work on this project is ongoing.
Microarray analysis of a crp regulatory mutant of Shewanella oneidensis
Amoolya Singh, University of California, Berkeley
Practicum Year: 2004
Practicum Supervisor: Jim Fredrickson, Laboratory Fellow, Biological Sciences Division, Pacific Northwest National Laboratory
This summer, I set out to learn the high-throughput experimental techniques that generate the biological data feeding my (theoretical) thesis work, and to study in detail the gene regulatory networks activated by environmental stress in various ecological niches. I focused on two strains of the model bacterium Shewanella oneidensis MR-1, one wild type and the other a crp regulatory mutant which, in addition to being defective in catabolite repression and carbon source selection, is also deficient in anaerobic respiration (Saffarini, Schultz et al. 2003). Both strains had been cultured in chemostats at PNNL in August 2003 with a variety of nutrient sources, electron acceptors, and oxygen concentrations.
Computer Simulations of The Active Site of Ras/Raf
Seung Lee, Massachusetts Institute of Technology
Practicum Year: 2003
Practicum Supervisor: David A. Dixon, Associate Director, Theory, Modeling and Simulation, Pacific Northwest National Laboratory
The Ras proteins are a family of the GTP-activated molecular switches that control signaling pathways for gene expression, cell proliferation and differentiation. The GDP-bound OFF (inactive) state can be activated by exchanging GDP with GTP. The GTP-bound ON (active) state can be turned off by hydrolysis of GTP to GDP. This hydrolysis rate can be increased by factors of up to 105 with the presence of GTPase activating proteins (GAPs). Mutations were found on Gln-61 of Ras proteins in 30% of human tumors. These mutations are characterized by decrease in hydrolysis rate and insensitivity to GAP. Ras also interacts with the well characterized effector protein Raf-1 to trigger the pathways for gene expression. We turn to numerical analyses, molecular dynamics and quantum mechanics, to elucidate the specific chemical interactions of these protein-protein complexes.
Evaluation of a Variety of Chemistry-transport Coupling Methods for Geochemical Groundwater Transport
Christopher Gesh, Texas A&M University
Practicum Year: 1995
Practicum Supervisor: Dr. Steven Yabusaki, , Environmental Research Center, Pacific Northwest National Laboratory
I helped evaluate the relative effectiveness of three methods for coupling the chemistry and transport code. I wrote several one dimensional codes for testing purposes and modified an existing three dimensional code. The methods we considered were global implicit, operator splitting and sequential iteration.
Odor Classification with Fuzzy ARTMAP
Lars Liden, Boston University
Practicum Year: 1995
Practicum Supervisor: Dr. Paul Keller, , Environmental Molecular Science Laboratory, Pacific Northwest National Laboratory
The ability of an artificial neural network known as Fuzzy ARTMAP to classify and identify odors present in an enclosed environment was examined. Data collected from an array of widely-tuned tin-oxide sensors formed a distributed representation (or signature) for chemicals present in an enclosed environment. The Fuzzy ARTMAP network learned to correctly classify sensor activations into categories and identify which chemical or combination of chemicals was present.
A Parallel Implementation of the Car-Parrinello Algorithm
James Wiggs, University of Washington
Practicum Year: 1992
Practicum Supervisor: Dr. John Lafemina, , Materials and Interfaces, Pacific Northwest National Laboratory
The Car-Parrinello algorithm is a computationally expensive technique for doing ab-initio molecular dynamics and global energy minimization of a coupled ionic-electronic system. We have developed a portable parallel implementation of the algorithm to be used on such machines as the iPSC/860, the Intel Delta or Paragon, the CM-5, etc.