Computer Science Colloquium Series Fall 2018

GPU Performance of the E3SM Spectral Element Atmosphere Dynamical Core

Dr. Mark Taylor, Chief Computational Scientist, Computational Science, Sandia National Laboratories

Wednesday, November 14, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

I will give an overview of the Department of Energy's "Energy Exascale Earth System Model" (E3SM), including Sandia’s role in numerical algorithms, parallel scalability and computational performance. E3SM is designed to run on upcoming next generation DOE supercomputers. Adapting simulation codes to these new architectures is expected to be more disruptive than the previous transition from vector to massively parallel supercomputers. E3SM development is driven by several grand challenge science questions focused Earth's cryosphere, biogeochemical and water cycle systems. E3SM has a new land and atmosphere component models branched from the CESM v1.2, coupled to new MPAS ocean, sea ice and land ice models.

The E3SM atmosphere component model's dynamical core is based on the spectral element method. I'll present performance results from our Fortran and C++/kokkos implementations of this dynamical core on a range of processors (Ivy Bridge, Haswell, Skylake, KNL, and P100 and V100 GPUs). With a careful implementation, the C++ code is competitive with the Fortran code on all processors and also supports GPUs.

This work allows us to evaluate the effectiveness of GPU architectures for modern climate models. After normalizing for power consumptions, we see that with sufficient work per node, the v100 GPU can obtain significant speedups over conventional Xeons. But in the strong scaling limit, we see little or no improvement over conventional architectures. Unfortunately, climate models are run close to their scaling limit in order to meet throughput requirements, and thus porting to GPUs is not cost effective. Instead, GPU systems should be used in other regimes. Two regimes where GPU systems can provide a real benefit are super-parameterization and ultra high resolution simulations running at throughput rates suitable for short process studies.

Bio:

Dr. Mark Taylor is a mathematician who specializes in numerical methods for parallel computing and geophysical flows. He currently serves as Chief Computational Scientist for the DOE's Energy Exascale Earth System Model (E3SM) project. He developed the mimetic/conservative formulation of the spectral element method used in E3SM's atmospheric component model. Mark received his Ph.D. from New York University's Courant Institute of Mathematical Sciences in 1992. He joined Sandia National Laboratories in 2004 and was promoted to Distinguished Member of the Technical Staff in 2018. In 2014 he was awarded (with Drs. David Bader and William Collins) The Secretary of Energy Achievement Award for his work unifying the Department of Energy's climate modeling research community, enabling the development of high-resolution fully-coupled climate-system simulations. Previously Mark served as co-chair of the Community Earth System Model's Atmosphere Model Working group (2011-2014) and as a member of the Korean Institute of Atmospheric Prediction Systems' Science Advisory Committee (2013-2016).

Communication Scheduling for Energy-Aware Computation

Dr. Varsha Dani, Department of Computer Science, The University of New Mexico

Wednesday, November 7, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Imagine a network of sensors spread throughout a forest, whose job is to detect and warn about forest fires.  When signs of a fire are detected, the sensors communicate with each other by sending radio signals.  The goal is to spread this information throughout the network as quickly as possible.  To make such a scheme economical, it is important that these sensors be very energy-efficient.  In this setting, communication costs are the single greatest drain on the devices' batteries. Unfortunately, for small devices communicating across relatively short distances, it costs nearly as much energy to listen for possible radio signals as it does to send them.  Consequently, instead of the more traditional way of measuring communication cost in terms of how many bits are sent, we also count the total amount of time spent sending or listening for messages.  Further difficulties are caused by not knowing in advance which devices are close enough to hear one another, and by the possibility of collisions, where messages are lost because two or more devices try to transmit at the same time. We will discuss some new algorithms for enabling one-to-all broadcasts in this setting, and of related problems, such as energy-efficient breadth-first search. 

Joint work with Yi-Jun Chang and Seth Pettie, of U. Michigan, Thomas Hayes, of UNM, and Qizheng He and Wenzheng Li, of Tsinghua U.

Bio:

Dr. Varsha Dani received her Ph.D. in Computer Science at the University of Chicago.  Her research interests include algorithm design and analysis, distributed computing, and communication, among other topics in theoretical computer science. 

Sparse Matrix-Matrix Multiplication: An MPI+X Story

Christopher Siefert, Principal R&D Staff member, Sandia National Laboratories

Wednesday, October 31, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

In high performance computing, one often hears about "MPI+X", which usually indicates the use of distributed memory for off-node parallelism (MPI) and shared-memory for on-node parallelism (X).  This talk will look at sparse matrix-matrix multiplication (SpGEMM), a computational kernel used in algebraic multigrid-based linear solvers for partial differential equations. For distributed memory, we will present an O(1) communication algorithm for neighbor discovery in the context of SpGEMM. With respect to OpenMP shared memory, we will present a thread-parallel algorithm which outperforms standard libraries on conventional architectures.  Finally, we will present some early work on "fused" OpenMP kernels, again motivated by algebraic multigrid.  Here we will demonstrate that the interfaces in standard libraries can be a poor fit for MPI+X applications and argue that different interfaces (and associated algorithms) could improve performance.

Bio:

Dr. Christopher Siefert received his Ph.D. in Computer Science with the Computational Science and Engineering option from the University of Illinois at Urbana-Champaign in 2006.  He is currently a Principal R&D Staff member in the Scalable Algorithms department at Sandia National Laboratories.  His research interests include numerical linear algebra, high performance computing, machine learning and computational electromagnetics.

Making Repairs in Description Logics More Gentle

Franz Baader, Professor, Dresden University of Technology

Wednesday, October 24, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Description Logics are used to define ontologies in many application areas. As the size of DL-based ontologies grows, tools that support correcting errors in such ontologies become more important. The classical approach for repairing a Description Logic ontology in the sense of removing an unwanted consequence is to delete a minimal number of axioms such that the resulting ontology no longer has this consequence. However, the complete deletion of axioms may be too rough, i.e., it may also remove consequences that are actually wanted. To alleviate this problem, we propose a more gentle notion of repair in which axioms are not deleted, but only weakened. On the one hand, we investigate general properties of this gentle repair method. On the other hand, we propose and analyze concrete approaches for weakening axioms expressed in the Description Logic EL. Though being quite inexpressive, EL has been used to define biomedical ontologies, such as the large medical ontology SNOMED CT.

(This talk reports on joint work with Francesco Kriegel, Adrian Nuradiansyah, and Rafael Peñaloza.)

Bio:

Prof. Franz Baader (15 June 1959, Spalt) is a German computer scientist at Dresden University of Technology. He received his PhD in Computer Science in 1989 from the University of Erlangen-Nuremberg, Germany, where he was a teaching and research assistant for 4 years. In 1989, he went to the German Research Centre for Artificial Intelligence (DFKI) as a senior researcher and project leader. In 1993 he became associate professor for computer science at RWTH Aachen, and in 2002 full professor for computer science at TU Dresden. (Bio taken from Wikipedia)

Correct Data-Plane SDN Programming

Jedidiah McClurg, Assistant Professor of Computer Science, The University of New Mexico

Wednesday, October 17, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

As network switching hardware becomes faster and more programmable, network functionality that was once confined to the controller is being pushed into the data-plane. For example, recent work on adaptive congestion control, heavy hitter detection, etc., has utilized stateful switches to execute the desired application with only minor (or zero) controller involvement. Well-known results show that managing global state in this distributed context is complicated, due to the tension between correctness (consistent views of global state) and efficiency (high throughput), and previous approaches for data-plane SDN programming provide little to no built-in support for addressing this difficulty. In this talk, Jedidiah proposes a new approach, featuring a general and intuitive network programming language which allows operators to write correct-by-construction data-plane programs with global state, and a compiler which in turn produces efficient executable code to run on modern SDN switches. Additionally, Jedidiah discusses a program synthesis-based approach for correctly composing multiple network functions into a single data-plane program.

Bio:

Jedidiah McClurg is new faculty in Computer Science at The University of New Mexico. He received his Ph.D. from the CUPLV group at the University of Colorado Boulder in 2018. He is currently working on research in synthesis and verification of software-defined network (SDN) programs, but has broad interest in programming languages, formal methods, and networking. His overall goal is to develop tools and techniques to help programmers write better code.

Data exploration and analytics: fundamental algorithms and emerging applications

Gautam Das, Distinguished University Chair Professor of Computer Science and Engineering, University of Texas at Arlington

Wednesday, October 10, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

In recent years, there have been many important advances in data exploration and data analytics techniques, both in fundamental algorithms as well as novel application domains. Our own research has focused on fundamental algorithms for data exploration such as query results ranking and top-k retrieval, as well as algorithms for data analytics such as sampling-based aggregate estimation. We have also made significant efforts to apply our algorithmic techniques to a variety of emerging application domains such as social media and the deep web, where the problems are challenging because of limiting data access interfaces. Our talk will focus on describing the interesting technical challenges and algorithmic innovations necessary to address these emerging problems, as well as the challenges and opportunities ahead.

Bio:

Gautam Das is the Distinguished University Chair Professor of Computer Science and Engineering, Director of the Database Exploration Laboratory (DBXLAB), and Director of the Big Data Analytics Center (BigDAC) at the University of Texas at Arlington. He graduated with a B.Tech in computer science from IIT Kanpur, India, and with a Ph.D in computer science from the University of Wisconsin, Madison. Dr. Das has broad research interests in all aspects of Big Data Exploration, including databases, data analytics and mining, information retrieval, and algorithms. He has published over 200 papers, with many of them appearing in premier conferences and journals. His work has received several awards, including the IEEE ICDE 10-Year Influential Paper award received in 2012. He has been on the Editorial Board of the journals ACM TODS and IEEE TKDE, as well as well as in program committees of numerous conferences, including serving as General Chair of ACM SIGMOD/PODS 2018. Dr. Das's research has been supported by grants from National Science Foundation, Army Research Office, Office of Naval Research, Department of Education, Texas Higher Education Coordinating Board, Qatar Foundation, AT&T, Microsoft Research, Nokia Research, Cadence Design Systems and Apollo Data Technologies.

Emergent Social Structures Arising Online in Disaster

Marina Kogan, Assistant Professor of Computer Science, The University of New Mexico

Wednesday, October 3, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Natural disasters are associated with breakdown of existing structures, but they also result in creation of new social ties in the process of self-organization and problem solving by those affected. This talk focuses on emergent social structures that arise in the context of crisis. Specifically, it considers collaborative work practices, social network structures, and organizational forms that emerge on social media during disasters. Social media platforms support highly-distributed social environments, and the social structures that emerge in these contexts are affected by the affordances of their technical features. Specifically, this talk investigates the social structures that emerge during disasters in three social media activities: retweeting, crisis mapping in OpenStreetMap (OSM), and Twitter reply conversations. Finally, the talk positions the findings from the three studies within the larger context of high-tempo, high-volume social media activity and highlights how the proposed framework reveals larger patterns within the social structures across contexts.

Bio:

Marina Kogan is an Assistant Professor of Computer Science at The University of New Mexico. Her research examines how people coordinate and problem-solve in crisis via social media interaction. Kogan received a Bachelor’s degree in Computer Science and another in Sociology at the City University of New York, followed by an MA in Sociology at the University of Illinois at Urbana-Champaign. With background in both computer and social science, Kogan crosses disciplinary boundaries in her work in social computing and human-centered data science, where she applies and develops methods that both harness the power of computational techniques and account for highly situated nature of the social activity in crisis.

Decomposition of aggregate networks provides insight into cognitive phenomena

Nicole Beckage, assistant researcher, University of Wisconsin

Wednesday, September 26, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

I will examine how procedures for optimally searching through "multiplex" networks (networks made of multiple simple graphs) capture human learning and search patterns. Prior work on semantic memory (people's memory for facts and concepts) has primarily focused on modeling similarity judgments of pairs of words as distances between points in a high-dimensional space (e.g., LSA by Landauer et al, 1998; Word2Vec by Mikolov et al. 2013). While these decisions seem to accurately account for human similarity judgments in some contexts, it is very difficult to interpret high dimensional spaces, making it hard to use such representations for scientific research. Further, it is difficult to adapt these spaces to a specific context or task. Instead, I define a series of simple networks that construct a multiplex network, where each network in the multiplex captures a "sense" or type of similarity between items -- for example the relationship between "moon" and "lamp" as compared to "moon" and "ball". I then optimize the "influence" of each of these feature networks within the multiplex framework where the weight of each network corresponds to the importance of each relationship. I use this approach to investigate how humans acquire language, and search through semantic memory. The resulting weighting of the multiplex can capture human attention and contextual information in these diverse domains. I explore how this approach can provide interpretability to multi-relational data and provide new insights in psychology and other fields by developing an optimization framework that considers not only the presence or absence of relationships but also the nature and importance of the relationships.

Bio:

Nicole Beckage is an assistant researcher at the University of Wisconsin in the Department of Psychology. She was previously an assistant professor in the Department of Electrical Engineering and Computer Science. She has dual PhDs from the University of Colorado, Boulder in Computer Science and Cognitive Science. Her work focuses on the use of machine learning and data science to build interpretable models to provide insight into human cognition, decision making and learning.

Quantum Computation and Optimization

Elizabeth Crosson, Assistant Professor, Center for Quantum Information and Control, The University of New Mexico

Wednesday, September 19, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Can quantum computers solve optimization problems much more quickly than classical computers? One major piece of evidence for this proposition has been the fact that quantum annealing (QA) can find the minimum of some cost functions exponentially more quickly than local search algorithms like classical simulated annealing. In this talk I'll describe a classical Markov chain Monte Carlo algorithm inspired by QA that can also exponentially outperform simulated annealing and classical local search, demonstrating that quantum algorithms can also lead to new methods in classical optimization. Along the way I will introduce the notion of quantum constraint satisfaction problems and their role in quantum computation, starting with the definition of a qubit and assuming only a background in linear algebra.

Bio:

Elizabeth Crosson has done research in many areas related to the theory of quantum computation, including quantum algorithms, quantum fault-tolerance, quantum computational complexity, and classical simulations of quantum systems. Prior to joining UNM Physics and the Center for Quantum Information and Control, she completed her Physics PhD at the University of Washington in 2015 and was a postdoc at the Caltech Institute for Quantum Information and Matter.

Scientific Machine Learning for Subsurface Problems

Youzuo Lin, staff scientist, Los Alamos National Laboratory Geophysics Group

Wednesday, September 12, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Subsurface is full of complexity and uncertainty. Current subsurface exploration methods can be expensive and yield limited accuracy. Machine learning has been used as a powerful tool in data analysis. While machine learning produces unprecedented success in conventional AI tasks, their applicability in scientific analysis problems including subsurface is even more challenging and exciting. This encompasses a wide range of problems including solving inverse problems, monitoring geologic formation changes due to fluid injection, detecting small but useful signatures out of large datasets, etc. Comparing to conventional AI domains, in scientific domains such as subsurface there are several unique challenges: restrictions in the measurement process, lack of availability of annotated data, and need for domain knowledge. I will demonstrate through some examples how we address those challenges.

My talk consists of two parts. In the first part of my talk, I will present novel approaches to characterize the subsurface using geophysical datasets. In particular, I will demonstrate how machine-learning techniques can be used to solve existing challenging problems such as geologic fault detection and full-waveform inversion. Various machine learning techniques will be covered including paired-image translation, data reduction using randomization, and spatial-temporal recurrent neural networks. In the second part of my talk, I will also quickly go through a few other promising examples of employing machine learning to subsurface problems such as monitoring geologic formation changes and detecting induced earthquakes. Through this talk, I will share both the successful and some not-so-successful results to demonstrate the potential and challenges of using machine learning approaches to solving existing problems in the subsurface.

Bio:

Youzuo Lin is a staff scientist in the Geophysics Group at Los Alamos National Laboratory. His research interests lie in computation and machine learning methods, and their applications in various geoscience problems including carbon storage, geothermal exploration, groundwater modeling, global seismology, and remote sensing imagery analysis. Youzuo received his Ph.D. in Applied Mathematics from Arizona State University in 2010. After completing his Ph.D., he was a postdoctoral fellow in the Geophysics Group at Los Alamos National Laboratory from 2010 to 2014, and then converted as a staff scientist.

A Volumetric Deep Learning Model for Rapid and Accurate Semantic Segmentation of Brain Imaging Data

Alex Fedorov, Ph.D. student, Electrical & Computer Engineering, and graduate research assistant, The Mind Research Network, The University of New Mexico

Wednesday, September 5, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Large-scale brain imaging studies of health and disease require processing and analysis of extensive multi-subject datasets. Complexity increases even further when segmenting structural MRI of the brain into semantically defined regions (such as anatomical areas). Current automatic approaches are time-consuming and hardly scalable; they often involve many error prone intermediate steps and don't easily utilize other available modalities. To alleviate these problems, we propose an automatic end-to-end segmentation model that is conveniently trained on the output produced by the state of the art (e.g. high-dimensional surface-based) models and can be directly applied to brain imaging data requiring no preprocessing. The model is a deep volumetric fully convolutional neural network with dilated filters that produce stronger regularization and use fewer parameters than the classical approach. The resulting segmentation is more than two orders of magnitudes (from more than 10 hours to under 2 minutes) faster than the state of the art and also more accurate. Importantly, segmentations produced by the model are deterministic: the same inputs lead to the same results. We evaluate the model's performance on over 1300 subjects and compare it with the state of the art on a large normative dataset as well as one identifying differences between schizophrenia patients and healthy controls. Results provide strong evidence that the decrease in processing speed did not harm the prediction accuracy. We also show an increase in the overall accuracy when multi-modal datasets are utilized (not an easy task for typically used models). Our model can expand to large-scale studies and can be useful in clinical scenarios, by reducing the delay between the patient screening and the result. Our approach will also simplify and encourage sharing of models, as the size of a fully trained model is under one megabyte, again three orders of magnitude improvement over the models in its class.

Bio:

Alex Fedorov is a Ph.D. student at The University of New Mexico, Electrical & Computer Engineering Department and Graduate Research Assistant at The Mind Research Network under the supervision of Prof. Vince D. Calhoun and Dr. Sergey Plis. His research is focusing on developing Deep Learning methods with applications to Neuroimaging. This summer he has interned at Mila, one of the highest ranked research labs for Machine Learning and Deep Learning, where he has worked on Representation Learning under the supervision of Prof. Devon Hjelm and Prof. Yoshua Bengio.

On Oblivious Filtering of Data Streams

Abhinav Aggarwal, Ph.D. student, Computer Science, The University of New Mexico

Wednesday, August 29, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

In today's highly data-driven world, computing on encrypted databases while hiding access patterns from an observant adversary is a necessity to provide the required privacy and security guarantees. While the existing solutions claim optimal protection against information leakage in an offline setting (where the size of the database is small), a formal analysis is missing to back these claims. Moreover, it is not entirely clear if these results can directly extend to the online setting (where look ahead and multiple passes are not (allowed) while preserving similar guarantees.

Motivated by the formal framework developed to study timing channel attacks on trusted devices and sensitive sources of information, we develop an information theoretic framework to qualitatively analyze the privacy guarantees of algorithms that wish to perform required tasks on data streams. I will focus our attention to only the filtering operation in this talk and extend the existing work to an online setting, where only a single pass over the entire database or the incoming data stream is allowed. We establish lower bounds on the memory requirements and the information leakage for the class of oblivious filtering operations that allow the adversary to infer the memory size at specific times of the protocol. Our analysis shows that with a memory of size B, any such protocol must leak of the order of (log B)/B bits of information per element. While most existing work as well as our solution assume that the memory used by the system runs inside an (expensive) trusted execution environments, we show how the use of an ORAM can help reduce the size of trusted memory by carefully accessing unprotected memory.

Our analysis shows that when the size T of input elements is more than log the size M of the unprotected memory, then we only need a trusted memory of size M/T log (M/T) < M. This implies that for a stream containing a million positive elements of 1Mb size each, the trusted memory has reduced from 1Tb (as required to store all elements in the trusted memory) to about 20Mb, which is more than 51000x reduction. Furthermore, if we allow w-window order preservation (order preservation across windows but not necessarily within), then we only need a trusted memory of size M/(wT) log (M/(wT)), implying that for the example above, a window of size 16 reduces the trusted memory requirement to only about 1Mb -- further 20x reduction in size. Combine this with 2x reduction in delay for every unit reduction in the buffer size, this implies almost 2 \times 10^6 times faster outputs.

Bio:

Abhinav Aggarwal is a Ph.D. candidate at The University of New Mexico Computer Science Department, working with Prof. Jared Saia on robust interactive communication protocols and resource competitive analysis. He likes working on interesting mathematically challenging problems, mainly at the intersection of security and theoretical aspects of distributed computing. He has interned with several companies like Microsoft, Google and Visa Research for his projects that span across different aspects of secure and fault-tolerant distributed systems. He has also served on various graduate student organizations on the campus, including CSGSA, GPSA, and ISA.

Towards Personalized Conversational AI

Prof. Yun-Nung (Vivian) Chen, Assistant Professor, Carnegie Mellon University, National Taiwan University

Wednesday, August 22, 2018
Centennial Engineering Center 1041
2:00-3:00 PM

Abstract:

Interacting with machines via natural language has been an emerging trend. The goal of developing open-domain dialogue systems that not only emulate human conversation but fulfill complex tasks, such as travel planning, seemed elusive. Recent advances in deep learning enabled new research frontiers for end-to-end goal-oriented conversational systems. This talk will review the research work about end-to-end situated dialogue systems, with components of situated language understanding, dialogue state tracking, dialogue policy, and language generation. In the dialogue framework formulated as a collaborative game, the conversational agent can be bootstrapped using user simulation and then further improved through interaction with real users.

Bio:

Prof. Yun-Nung (Vivian) Chen is currently an assistant professor at the Carnegie Mellon University, National Taiwan University. She earned her Ph.D. degree from , where her research interests focus on spoken dialogue system, language understanding, natural language processing, and multi-modal speech application. She received MOST Young Scholar Fellowship 2018, Google Faculty Award 2016, two Student Best Paper Awards from IEEE SLT 2010 and IEEE ASRU 2013, a Student Best Paper Nominee from Interspeech 2012, and the Distinguished Master Thesis Award from ACLCLP. Prior to joining National Taiwan University, she worked in the Deep Learning Technology Center at Microsoft Research Redmond.