diComputing/CORD Presentations – ORION THINK Conference

Compute Ontario supports the development of preeminent, provincial advanced computing, including high-performance computing (HPC), and regularly gives that community the opportunity to learn about each other’s work.

As part of Compute Ontario’s annual Research Day and the diTHINK diComputing track, we have received abstracts for contributed and poster presentations from professors, postdoctoral, graduate, and undergraduate students.

Contributed Presentations

Contributed presentations will be in the form of a 15-minute, oral, slide-based presentation, followed by two minutes of questions. These presentations will occur in morning and afternoon contributed presentation tracks.

Morning presentations

10:30am – James Chow, Medical Physicist and Assistant Professor, University of Toronto

Evaluation of Radiation Treatment Plan using Big Data and Cloud Computing

Abstract: To compare the quality of radiation treatment plans from different institutions worldwide using their specific treatment planning systems (TPSs), treatment units and radiation delivery techniques, we propose to build a dynamic big database to facilitate plan evaluation and comparison. This big database is less concerned with how to create the treatment plan than the dosimetric results in the form of dose volume histogram (DVH). A web-based interface is linked to the cloud and contains different treatment sites such as head-and-neck, lung, breast and prostate.

In each site, the user uploads the DVH of the plan and data processing such as Gaussian error function modelling to be carried out on the cloud. Dosimetric and radiobiological parameters for plan evaluation will be calculated and sent back to the user, together with a qualitative comparison based on the big data. This on-growing big database can help planners from different institutions for plan evaluation.

10:50am – Graham Taylor, Assistant Professor, University of Guelph

Hardware Accelerators for Deep Learning

Abstract: Deep learning is a branch of machine learning which is based on learning feature hierarchies from highdimensional, complex data sets. It has transformed industry, powering the major players such as Google, Facebook, IBM and Microsoft, as well as hundreds of startups, enabling new products and services in areas such as computer vision, speech analysis, and natural language processing. It has ridden the wave of cheap and widely available computation (namely general-purpose GPUs) and large human-annotated datasets.

In this talk, I will highlight some of our group’s recent efforts in using hardware accelerators to speed up deep learning algorithms. First, I will motivate the need for hardware accelerators from a model search perspective. Then I will describe multi-GPU implementations of convolutional neural networks. Finally, I will describe an implementation of convnets on Field-programmable gate arrays (FPGAs), which are a type of low-power, reconfigurable hardware device. FPGAs can achieve comparable throughput to GPUs, but at about an order of magnitude less power.

11:10am – David Hilts, Morouney Rusu, and Robert Vaughan, Students, Wilfrid Laurier University

Monte Carlo AIXI Parallel MPI Approximation

Abstract: The reinforcement learning area of machine learning studies how agents in environments should behave when the agent receives feedback after each action it take. AIXI is a mathematical solution to the general reinforcement learning problem which performs optimally, unfortunately AIXI is not computable, thus several approximations have been made.

Our contribution is a parallelization of the learning algorithm described in “A Monte-Carlo AIXI Approximation” (Joel Veness et al.) which uses Context Tree Weighting and Monte Carlo tree search to approximate an AIXI agent. To approximate this non-computable result, we investigate a parallel method using MPI to approach and demonstrate results on the SHARCNET computing cluster.

Afternoon presentations

2:45pm – Shimiao Zhang, M.A.Sc Candidate, McMaster University

Precursor Effects on the Structure and Properties of Polymer Networks Synthesized using Molecular Dynamics

Abstract: The effects of precursor topology on the formation, structure and mechanical properties of polymer networks are studied using coarse-grained molecular models. Cross-linked polymer networks are synthesized with molecular dynamics from three different sets of molecular precursors with varying chain length and their structure and properties are compared.

Little difference is observed between these networks in the radial distribution function, macroscopic statistics of network connectivity, and glass transition behaviors. The elastic modulus of the network is found to correlate strongly with the number of elastic strands in the network, except at the highly-crosslinked limit where substantial discrepancy is observed between networks from different precursors. Although these final networks contain a similar level of structural defects, the choice of precursor has a significant effect on the spatial distribution of the defects, which explains the precursor dependence of their mechanical property observed in the tensile test.

3:05pm – Jonathan Gillett, Graduate Researcher, University of Ontario Institute of Technology

An Analysis of Bitcoin, and Its Impact on Global Finance

Abstract: Bitcoin is a novel digital currency that is distributed using a decentralized peer-to-peer network. It also uses cryptography, and a compute-intensive process referred to as “mining” to ensure its scalability and security. This presentation provides a detailed overview of Bitcoin, including important aspects such as the decentralized peer-to-peer network; the deeply intertwined role of cryptography; the blockchain, which provides a distributed ledger of all transactions; and lastly, the mining process, which incentivizes users with Bitcoin and secures transactions on the network.

Following this is a discussion of our research, which involves the analysis of hundreds of millions of transactions on the blockchain; the daily trading activity and economic impact of the Bitcoin network; and lastly our metrics, which assist in understanding the network and its impact on global finance.

3:25pm – Erik Schnetter, Research Technologies Group Lead, Perimeter Institute for Theoretical Physics

Using the Julia Programming Language for High-performance Computing

Abstract: Julia is a new programming language designed particularly for technical computing. Julia contains features from Fortran, C++, Python, and Mathematica that make it particularly interesting for tasks in numerical simulations, data analysis, or visualization.

I will give a whirlwind introduction to Julia and some of the features that make Julia especially interesting, and will highlight with an example how Julia can be used in high-performance computing. All code shown will be available for download as Jupyter (IPython) notebook.

Poster Presentations

Poster presentations will be organized during the morning and afternoon conference breaks, lunch, and the afternoon networking session.

We anticipate to offer prizes for best student contributed presentation and best student poster presentation.

Evaluation of Radiation Treatment Plan using Big Data and Cloud Computing by James Chow, University of Toronto

Effects of drag-reducing polymers on the transition and growth of turbulent coherent structures by Xue Bai, McMaster University

Drag-reducing polymers have profound impact on the dynamics of turbulence, not only in the statistically-converged stage but also in its transient development process. Our interest in this area is motivated by two prominent unsolved problems in viscoelastic turbulence: maximum drag reduction and the mechanisms of transition to turbulence in viscoelastic fluids.

The pivotal solution object controlling the laminar-turbulent transition, the so-called “edge state,” is numerically computed. Direct numerical simulation is then used to track the transient development of these marginally turbulent states into fully developed turbulence. The transition process of Newtonian fluids is characterized by a strong breakdown event where turbulent structures sharply intensify across the channel before their eventual decay into the typical coherent structures in the turbulent basin. Polymer additives are found to effectively suppress these high-intensity vortices, which leads to a different transition pathway where a strong breakdown is bypassed.

Web-based Information Platform on Ethnic Cultural Practices in Cancer Treatment by James Chow, University of Toronto and Kay Li, York University

Abstract: We are building an information platform for Asian Canadian community using the cloud computing technology to provide medical and scientific validity of various cultural practices in cancer treatments.

The database of the platform contains different Asian ethnic cultural practices such as Chinese qigong, Hindu Yoga, Japanese shiatsu massage and India aromatherapy. The platform provides personalized programs to the users through a web-based graphical user interface. Resources from different cultural practices will be optimized by the Monte Carlo method, which will be performed on the cloud. The
platform will explore how various ethnic culture practices can be used to complement cancer treatments, in particular to alleviate side effects. Information in English and ethnic Asian languages is provided as there are Asian Canadian communities disadvantaged by language and cultural barriers.

Simulated Biophysical Experimental Techniques for Chlorhexidine in DMPC by Brad Van Oosten, Brock University

Abstract: We have investigated the use of molecular dynamic simulations and the MARTINI force field to simulate isothermal titration calorimetry and differential scanning calorimetry techniques. The goal of these simulations was to observe how well they can reproduce the concentration effects of the addition of the small molecule chlorhexidine into a model DMPC membrane.

We were able to mimic an isothermal titration calorimetry experiment by repeatedly adding a 1% concentration of chlorhexidine into a DMPC membrane. We observed the mechanism in which chlorhexidine enters the membrane, as well as some of the structural changes it causes on the membrane. We then performed a controlled cooling of the membrane to mimic a differential scanning calorimetry experiment. We then varied the concentration of chlorhexidine in order to observe the change of the lipid melting temperature due to the addition of chlorhexidine.

Alternative Representations of Julia Sets by Harold Hodgins, Wilfrid Laurier University

Abstract: Julia Sets are a way of visualizing the iterative properties of functions. Traditionally this is done using high resolution images but advances in 3D printing make it a viable option for creating tangible models. We will be presenting several 3D models to represent several properties iterative functions have.

Mathematical Modelling of Hepatitis C Virus: Quantifying the Contribution of Cell-free vs. Cell-to-cell Infection by Kenneth Blahut, Ryerson University

Abstract: Experiments have shown that the spread of hepatitis C virus (HCV) infections disseminate both distally via release and diffusion of cell-free virus through the medium, and locally via direct cell-to-cell infection. However, it is impossible to differentiate between the two modes of infection experimentally. Characterizing the relative contribution of infection for each mode has important implications for the control of HCV infections i.e. the selection of antivirals and/or cell receptor inhibitors used for treatment.

We have developed an agentbased computer model which explicitly incorporates both cell-free and cell-to-cell modes of infection. By using in vitro experimental data to constrain our model, we show that cell-to-cell infection is dominant, and can contribute up to 95% of the total infection.

Parallel Generation of Julia Sets by Scott King, Wilfrid Laurier University

Abstract: We have designed an algorithm to generate points (and colour magnitudes) for high resolution Julia set images using the base equation: fc(z) = z^2 + c. The coordinate generation is load balanced and is able to generate 15000×15000 pixel images with 1000 iterations deep.

Visualization Tool for Debugging Pilot Cluster Programs by Tianyi Bao, University of Guelph

Abstract: Pilot is an open source parallel programming library from the University of Guelph for C and FORTRAN cluster programs that aims to ease the learning curve for novice scientific programmers. Built as a thin layer on top of MPI, Pilot is called “A friendly face for MPI.” Until now, third-party visualization tools for MPI programs would display low-level operations confusing to Pilot programmers using the latter’s higher-level abstractions (such as channels and bundles).

After incorporating calls to MPI Parallel Environment (MPE) into the Pilot library, user programs can now automatically generate CLOG-2 log files suitable for visualization with portable, Java-based Jumpshot-4. Pilot API calls and even message data are displayed. In this way, programmers can understand the actual run-time message passing between processes more clearly, helping to diagnose and fix their logic. This feature will be available in the next public release of Pilot.

Identifying Neurodevelopmental Recessive Variants using Whole Exome Sequencing and Computational Autozygosity Mapping by Ricardo Harripaul, Centre for Addiction and Mental Health

Abstract: Intellectual Disability (ID) affects 1% of the population and overlaps with many psychiatric disorders such as Epilepsy, Schizophrenia, Bipolar Disorder and Autism Spectrum Disorder. We sequenced 181 families using the Illumina HiSeq 2500 and used a high performance cluster to perform computational autozygosity mapping and whole exome sequencing analysis, to identify variants associated with ID.

The use of high performance computing was instrumental in storing data, integrating different data types, analyzing data, and managing the large amount of data produced in this study. After alignment to the human genome and variant detection we identified 73 genes in these families and discovered 35 novel genes for ID including TRAPPC6B, MBOAT7 and SLAIN1. ID is extremely heterogeneous and our novel gene findings will translate into better diagnosis in the individualized medicine era where many sufferers go without a genetic cause.

Questions?

For inquiries regarding the scientific community at diTHINK, including contributing and poster presentations, please contact Dalibor Dvorski at ddvorski@conestogac.on.ca