Speaker Abstracts

To read the speakers abstract please click on the link below associated with their name.

Title: Data Centric Systems: Algorithm Exploitation & Evolving AI/Cognitive Examples

Speaker: Kirk Jordan, IBM Distinguished Engineer Data Centric Solutions IBM T.J. Watson Research & Chief Science Officer - IBM Research UK

Abstract: The volume, variety, velocity and veracity of data is pushing how we think about computer systems. IBM Research’s Data Centric Solutions organization has been developing systems that handle large data sets shortening time to solution. This group has created a data centric architecture initially delivered to the DoE labs at the end of 2017 and being completed in 2018. As various features to improve data handling now exist in these systems, we need to begin to rethink the algorithms and their implementations to exploit these features. This data centric view is also relevant for Artificial Intelligence (AI) and Machine Learning (ML). In this talk, I will describe the architecture and point out some of hardware and software features ready for exploitation. I will show how we are using these data centric AI/cognitive computing systems to address some problems in new ways as case studies.

Title: Why huge simulations are invaluable for understanding fluid flow physics

Speaker: Stephen De Bruyn Kops Department of Mechanical and Industrial Engineering UMass Amherst

Abstract: Many fluid flows are turbulent. We could not breathe without turbulence. In fact turbulence is a central mechanism in almost any sort of mixing so that things from simple combustion to weather and climate depend on turbulence. Thus the challenge in predicting a wide variety of flows comes down to predicting fluid turbulence, which is fundamentally three dimensional, non-linear, and multiscale in space and time. Because it is three dimensional and multiscale, turbulence measurements, either in-situ or in the laboratory, involve assumptions, and sometimes the same assumptions that scientists would really like to test. By the early 1970's, through, computers became sufficiently powerful to solve the partial differential equations that describe turbulent flow using a class of simulations called "direct numerical simulation" (DNS) that involve no models. By about the year 2000, DNS and laboratory experiments were seen as complimentary rather than the former being validated by the latter. Today, in some fields, DNS is used for studying flows with a far greater range of length and time scales than can be measured in the laboratory. Our largest DNS uses 3 trillion grid points and provides us a full description of the flow field as a function of space and time. A snapshot in time of the flow field is 48 terabytes. We can interrogate it to it to test virtually any theory or modeling or assumption that is proposed. We can show oceanographers, or combustion scientists, or climate modelers the validity of their assumptions in our simulated flows which, while extremely simple, involve ranges of length and time scales realistic for complex flows involving turbulence. We do not pretend to simulate the full complexity of, say, the ocean, but apply the full capability of modern computers to address the question of whether the assumptions built into in-situ measurements of the ocean are consistent with solutions of the governing equations with a realistic range of turbulence length and time scales.

Title: High-fidelity nuclear reactor simulations and the need for Exascale computing

Speaker: Benoit Forget, Department of Nuclear Science and Engineering, Massachusetts Institute of Technology

Abstract: This talk will introduce the ExaSMR project under the Exascale Computing Project (ECP) in support to the development of high fidelity simulations for nuclear reactor applications. The high cost of experimentation has complicated the development of novel nuclear technologies which enhances the need for high fidelity simulations. The ExaSMR project is a collaboration between Oak Ridge National Laboratory, Los Alamos Laboratory and MIT which aims to develop a modern approach to couple high fidelity Monte Carlo neutron simulations with detailed computational fluid dynamic models on modern computing architecture. This presentation will describe the complexity and dimensionality of the problem as it applies to small modular reactors (SMRs) and present various approaches that have been developed in improving the performance and fidelity of the simulations.

Title: Detecting Large Scale Chromosomal Rearrangements: A big data challenge

Speaker: Nurit Haspel Department of Computer Science UMass Boston

Abstract: Structural variations (SVs) are deletions, duplications and rearrangements of medium to large segments (>100 base pairs (bp)) of the genome. Such genomic mutations are often described as being the primary cause of many diseases, including cancer. Breakpoint detection using next generation sequencing (NGS) platforms is very challenging since computational methods to detect SVs face the big challenge of accurately predicting the precise location of breakpoints, which are typically spanned by a very small number of reads among hundreds of millions of DNA segments that are generated during the sequencing process. This is a classical Big Data problem.To address the problem we use a combination of dimensionality reduction, distance-preserving embedding and massive parallel computing, which allows us to solve the problem efficiently. Later, we use an SSE-multithread architecture to achieve an efficient search with higher accuracy.

Title: Workload Characterization Tools For Every Need: From Architecture Agnostic Classification of Communication to Trace-based Simulation of Multi-Threaded Workloads

Speaker: Mark Hempstead Department of Electrical and Computer Engineering, Tufts University Adjunct Professor, Department of Computer Science, Tufts University

Abstract: Workload characterization tools enable detailed instrumentation and analysis that help designers and researchers understand the features and characteristics of the workloads running on their systems. In this talk we describe a range of workload characterization tools that can classify communication in an architecture agnostic manner and detect workloads that could be accelerated by the same shared hardware.

Through communication classification, one can measure the amount and type of data that moves within a workload. This classification is not just counting the amount of data moved and computed, but also measuring whether the communication is unique or non-unique—the first time data has been seen or reuse of that data. Furthermore, the contextual locality of communication is just as important. That is, whether data has been referenced within an entity (path, function, or thread), or between entities. We have shown that this classification can be used for a diverse array of co-design research, including hardware accelerator design, thread mapping, hardware simulation, and design space exploration.

Current workload analysis methods are often embedded within a specific instrumentation tool (e.g. PIN, LLVM, Valgrind). Thus, when developers want to target a new hardware platform or use a new instrumentation front-end, the workload analysis must be rewritten. PRISM, a framework which enables workload analysis tools that are both developed and executed agnostic of the underlying hardware platform. The benefits of such a framework are that researchers can capture salient workload data, without domain specific knowledge of each hardware platform, and then reuse analysis tools across applications that use different hardware configurations.

Computer Architecture researchers want to be able to study multithreaded applications in simulation. Through communication classification we can represent architecture agnostic traces of high-performance computing workloads, SynchroTraces. These traces capture synchronization events and allow for the interleaving of threads to change during simulation. We demonstrate how SynchroTraceSim provides an order of magnitude improvement in simulation time over cycle accurate simulation.

Bio: Mark Hempstead is an Associate Professor in the Department of Electrical and Computer Engineering at Tufts University. His research group, the Tufts Computer Architecture Lab, investigates methods to increase energy efficiency across the boundaries of circuits, architecture, and systems. Currently, they are exploring the performance and energy benefits of heterogeneity in future many-accelerator architectures; the security implications of the thermal-side channel; methods to automatically generate shared hardware accelerators from source code; power-aware hardware and software for Android devices, and flexible workload characterization tools. His group has published innovations in several different research communities including high performance computer architecture, embedded systems, workload characterization, mobile-systems, and the Internet-of-Things (IoT).

Hempstead received a BS in Computer Engineering from Tufts University and his MS and PhD in Engineering from Harvard University working with Professors David Brooks and Gu-Yeon Wei. Prior to joining Tufts University in 2015 he was an Assistant Professor at Drexel University. He received the NSF CAREER award in 2014 and the Drexel College of Engineering Excellence in Research Award in 2014. He was honored for his achievements in teaching with the 2014 Drexel University Allen Rothwarf Award for Teaching Excellence given to one junior faculty member a year. He was the winner of the industry sponsored SRC student design contest in 2006 and Best Paper Nominee in HPCA 2012.

Title: Distributing Frank-Wolfe via Map-Reduce

Speaker: Stratis Ioannidis Department of Electrical and Computer Engineering Northeastern University

Abstract: Large-scale optimization problems abound in data mining and machine learning applications, and the computational challenges they pose are often addressed through parallelization. We identify structural properties under which a convex optimization problem can be massively parallelized via map-reduce operations using the Frank-Wolfe (FW) algorithm. The class of problems that can be tackled this way is quite broad and includes experimental design, AdaBoost, and projection to a convex hull. Implementing FW via map-reduce eases parallelization and deployment via commercial distributed computing frameworks. We demonstrate this by implementing FW over Spark, an engine for parallel data processing, and establish that parallelization through map-reduce yields significant performance improvements: we solve problems with 20 million variables using 350 cores in 79 minutes; the same operation takes 48 hours when executed serially.

This is joint work with Armin Moharrer.

Bio: Stratis Ioannidis is an assistant professor in the Electrical and Computer Engineering Department at Northeastern University, in Boston, MA, where he also holds a courtesy appointment with the College of Computer and Information Science. He received his B.Sc. (2002) in Electrical and Computer Engineering from the National Technical University of Athens, Greece, and his M.Sc. (2004) and Ph.D. (2009) in Computer Science from the University of Toronto, Canada. Prior to joining Northeastern, he was a research scientist at the Technicolor research centers in Paris, France, and Palo Alto, CA, as well as at Yahoo Labs in Sunnyvale, CA. He is the recipient of a Google Faculty Research Award and a best paper award at ACM ICN 2017. His research interests span machine learning, distributed systems, optimization, and privacy.

Title: Finding Genomic Elements That Are Extremely Conserved In Evolution Using Cache-Oblivious Computing

Speaker: Dmitry Korkin, Department of Bioinformatics and Computational Biology, Worcester Polytechnic Institute

Abstract: Modern genomics is driven by massive volume of data and stands at the forefront of high-performance scientific computing. Yet, we are still far from creating a complete encyclopedia of functional elements of the human genome. One of the reason lies in our incomplete understanding how genomes of the leaving organisms evolve. For instance, humans and mice diverged from their common ancestor about 75 million years ago, and ever since they started accumulating genetic changes in the genes they onces shared. Thus, even the most functionally important genes have some differences when we comparing different species. Yet, in 2004, an striking counter-example of this rule came about when two groups independently made an intriguing discovery. They found stretches of genomes that were evolutionary intact, sharing 100% identity between the genomes of human, mouse, and rat. The sequences were called ultraconserved elements (UCEs). In spite of the tremendous interest to the subject, the origins of ultraconservation are yet to be ascertained. Until recent, the existence of such extreme regions between more diverse species has not been known. One of the main bottlenecks was the string search algorithms which deemed not feasible for a diverse set of genomes or genomes that underwent drastic evolutionary changes. To tackle this problem, we have recently developed a principally new, cache-oblivious, approach to this problem. The idea of this approach is to develop a method that will be optimal from the algorithmic point of view as well as from the point of view of the computational hardware it is run on. Our method is designed to improve a standard hash mapping approach by optimizing the data exchange with the CPU’s cache memory. Compared to our original hash-mapping approach, the new cache-oblivious method achieves a remarkable speedup of up to 2,000 times, eliminating all previous computational bottlenecks. Most importantly, for the first time we discovered a large group of ancient ultraconserved elements that are conserved beyond tetrapods, suggesting that the origins of extreme conservation could be as early as 700 MYA. To the best of our knowledge, this is the first cache-oblivious approach in computational genomics, suggesting a paradigm shift in bioinformatics algorithms in order to deal with the exponentially growing data volume.

Title: Quantum Chemical Study of the Initial Self-Assembly of Aromatic Dipeptides Into Nanostructures for Biomedical Applications

Speaker: Maricris Mayes, Department of Chemistry and Biochemistry UMass Dartmouth

Abstract: Self-assembled peptide nanotubes have many potential applications in the fields of nanobiotechnology and nanomedicine. Peptides serve as excellent building blocks due to their high availability and versatility. Tyrosine-based dipeptides self-assemble to form higher order structures. In this presentation, we describe our work on linear and cyclic dityrosine (YY) and tryptophan-tyrosine (WY) to gain insights into the nature of intermolecular interactions contributing to the initial aggregation of dipeptide-based nanostructures. We use a variety of quantum-chemical methods with dispersion corrections and universal solvation model based on density in combination with energy decomposition and natural orbital bond (NBO) analyses. We show that hydrogen bonding is a major stabilizing force. The lowest energy structure for the linear YY dimer is characterized by Ocarboxyl∙∙∙H(O)tyr, whereas the lowest energy dimer of linear WY is dominated by Ocarboxyl∙∙∙H(N)trp and πtyr∙∙∙πtyr. The role of solvent is important as it impacts the strength and nature of interactions. The lowest energy for linear WY dimer in acetone is stabilized by Ocarboxyl∙∙∙H(O)tyr, πtrp∙∙∙H(C), and πtrp∙∙∙H(N). The thermodynamics of dimerization and stabilization energies of solvated dipeptides reveal that the dipeptide systems are more stable in the solvent phase than in gas-phase. NBO confirms increased magnitudes for donor-acceptor interaction for the solvated dipeptides.

Title: Electrostatic networks in natural enzymes: What can we learn for protein engineering?

Speaker: Mary Jo Ondrechen, Department of Chemistry & Chemical Biology Northeastern University

Abstract: Imagine a world where industrial chemical processes are catalyzed by enzymes, using less energy with fewer unwanted by-products than conventionally catalyzed or thermal processes. Because for most industrial chemical reactions there does not exist a natural enzyme that can serve as the catalyst, we must learn how to engineer new enzymes. Electrostatic interactions across networks of amino acids are important features that give natural enzymes their catalytic power. Understanding these networks is necessary to learn how to build these properties into in silico enzyme designs. Partial Order Optimum Likelihood (POOL) is a machine learning method developed by us to predict amino acids important for function, using the 3D structure of the query protein. POOL predicts networks of amino acids that impart the necessary electrostatic, proton-transfer, and ligand binding properties to the active residues in the first layer around the reacting substrate molecule(s). Typically these networks include first-, second-, and sometimes third- layer residues. POOL-predicted, multi-layer active sites with significant participation by distal residues have been verified experimentally by site-directed mutagenesis and kinetics assays for several different kinds of enzymes. An approach to build these properties into the initial designs for protein engineering is proposed. Acknowledgment: NSF MCB-1517290.

Bio: Mary Jo Ondrechen received an ACS-certified Bachelor’s degree in Chemistry from Reed College and the Ph.D. degree in Chemistry and Chemical Physics from Northwestern University in Illinois. After postdoctoral research appointments at the University of Chicago and at Tel-Aviv University in Israel, the latter as a NATO Postdoctoral Fellow, she joined the faculty at Northeastern University in Boston, Massachusetts. Currently she is Professor of Chemistry and Chemical Biology and is the Principal Investigator of the Computational Biology Research Group at Northeastern University. The focus of her research is on understanding enzyme catalysis, predicting the function of proteins from genomics, protein design, and the computational aspects of drug discovery. She is the past President of the Board of Directors of the North American Indian Center of Boston (NAICOB), has recently served on the Board of Advisers of the Interstate Technology and Regulatory Council (ITRC), and was the 2011-2013 Chair of the Board of Directors of the American Indian Science and Engineering Society (AISES). She is a co-PI on the 2014-2019 project “Lighting the Pathway to Faculty Careers for Natives in STEM,” an initiative to provide guidance and support to 100 Native STEM students who want to become faculty members at colleges, universities, and tribal colleges. She is also the Director of “Northeastern University Skills and Capacity for Inclusion (NU-SCI): Inclusive Excellence Catalyzed by Experiential Education,” funded by a grant from the Howard Hughes Medical Institute. At the 2018 American Chemical Society National Meeting in New Orleans she gave the Eminent Scientist Lecture). She is a proud member of Mohawk Nation.

Title: Using HPC for engine and fuel spray simulations

Speaker: Noah Van Dam Department of Mechanical Engineering UMass Lowell

Abstract: Internal combustion engines form the basis of the vast majority of the ground transportation fleet in the United States and across the world. Continuous research has developed modern engines that are cleaner and more fuel efficient than ever before. However, in support of a more sustainable future, these engines must continue to improve their efficiency and reduce their pollutant emissions while using non-traditional biofuels. Engine simulations are key to understanding how new biofuels may impact engine performance. But, for the results to be useful the numerical errors in the modeling must be understood. This talk will discuss two different engine modelling studies. The first uses a sensitivity analysis to quantify the effect of varying fuel properties such as heat of vaporization and density on overall engine performance. Experimentalists have tried quantifying this relationship previously, but a real fuel’s properties are inherently linked to the fuel’s chemical composition. Thus it is not possible experimentally to fully decouple the effects of a fuel’s properties such as heat of vaporization and its chemical composition. Simulations, however, have no such limitation, and are used to investigate the pure physical property effects. The second investigation is on the development of robust error estimates for stochastic Lagrangian-Eulerian spray models. These models have gained widespread use due to their relative computational efficiency, and have been validated against experimental spray measurements. However, a formal error analysis has yet to be fully developed due to the complex interaction of the Eulerian and Lagrangian descriptions of the fuel spray. Recently, some work has been done to try to begin to understand the contribution of statistical sampling error and local resolution errors on simple spray modeling simulations in order to build up a fuller error analysis that may be used to increase the accuracy of these methods for future engine simulations. This presentation will summarize that work, along with future directions for the verification of mixed Lagrangian-Eulerian methods.