REU Site Summer Research Program

• Dr. Nihat Altiparmak

  

  Faculty Profile

  View Dr. Nihat Altiparmak’s faculty profile

  Research Areas:

  • Operating systems
  • Data storage systems
  • Parallel and distributed systems

  Sample Project Description:

  Energy-Aware Optimization of Operating System Kernels for Ultra-Low Latency Storage Performance

  The operating system kernel is a complex piece of software that sits between the applications of a computer and its hardware, with the main goal of providing convenient access to computer hardware in an efficient and fair manner. From personal computers to servers and supercomputers, embedded systems and robots to mobile devices like phones and tablets, operating systems are heavily used to manage a variety of computer systems in today’s world.

  Since computer hardware is constantly improving and innovating, operating system kernels need continuous optimizations in order to maintain their efficiency. Today’s operating systems urgently require optimization in their storage stack, which manages access to data storage devices. Various innovations have taken place in the storage subsystem within the past few years, including wide adoption of the Non-Volatile Memory Express (NVMe) interface [1] and the emergence of new generation of Ultra-Low Latency (ULL) SSDs, which are broadly defined as providing sub-10 μs data access latency [2]. This new level of ULL storage device performance questions the suitability of existing kernel storage stack designs that are primarily optimized for older storage generations with higher data access latencies, such as flash-based SSDs and HDDs, with one and three orders of magnitude slower data access latencies, respectively.

  Efficiency of an operating system is generally measured in terms of its performance, where the energy impact of the optimizations is commonly neglected. In other words, energy efficiency is typically regarded as a second class citizen in operating system design. However, today’s data centers consume as much electricity as a city [3]. In addition, the energy efficiency of an operating system managing a battery-operated device such as a mobile phone or tablet can significantly affect the battery life of that device. Therefore, any optimization that is performed on an operating system kernel should be performed in an energy-aware manner.

  In this REU project, the undergraduate student(s) will investigate, analyze, and optimize the Linux kernel’s storage stack for ULL storage performance in an energy-aware manner. Each student will be assigned a specific storage stack section to investigate, such as submission, scheduling, or completion, and perform empirical experiments using real hardware: an Intel Optane SSD [4] (a type of ULL SSD) in a dedicated server with an Onset HOBO Power meter [5] to measure energy consumption. Using this experimental setup, students will first investigate and understand the working principles of their assigned storage stack layer. Next, they will experimentally analyze the efficiency of existing methodologies within their layers in an energy-aware manner, and make observations about their deficiencies. Finally, based on their observations, students are expected to propose and implement possible optimizations to eliminate these observed deficiencies.  Students will be trained on kernel development and will work in collaboration with other lab students experienced in operating system design and development.

  References Cited

  • NVM Express Base Specification, rev. 1.4. https://nvmexpress.org/wp-content/uploads/NVM-Express-1_4-2019.06.10-Ratified.pdf.
  • Bryan Harris and Nihat Altiparmak. Ultra-Low Latency SSDs’ Impact on Overall Energy Efficiency. In 12th USENIX Workshop on Hot Topics in Storage and File Systems (HotStorage 20). USENIX Association, July 2020.
  • Bryan Harris and Nihat Altiparmak. Monte Carlo Based Server Consolidation for Energy Efficient Cloud Data Centers. In 11th IEEE International Conference on Cloud Computing Technology and Science (CloudCom 2019), pages 263–270, Sydney, Australia, December 2019.
  • Intel Optane SSD 900P Series Product Brief. https://www.intel.com/content/dam/www/public/us/en/documents/product-briefs/optane-ssd-900p-brief.pdf.
  • Onset HOBO® UX120-018 Data Logger (datasheet). Onset. https://www.onsetcomp.com/datasheet/UX120-018.

• Dr. Antonio Badia

  

  Faculty Profile

  View Dr. Antonio Badia’s faculty profile

  Research Areas:

  • Databases
  • Data Science
  • Artificial Intelligence

  Sample Project Description:

  KNOWLEDGE BASE ANALYSIS

  Knowledge Bases (KBs) store statements about the world in a format that can be used for reasoning, like RDF. Moreover, KBs organize all their information into hierarchies, relying on some taxonomy or ontology. YAGO is an example of modern Knowledge Base, with facts extracted from Wikipedia and organized using the schemas from WordNet; it has recently been expanded to use schema.org. YAGO can be downloaded from http://www.yago-knowledge.org; the YAGO taxonomy is stored in a separate file.

  The purpose of this project is to analyze YAGO’s taxonomy. Several steps are involved: first, we want to create a file with simplified format, i.e. to go from

  <http://yago-knowledge.org/resource/Germany>    <http://www.w3.org/1999/02/22-rd

  f-syntax-ns#type>       <http://schema.org/Country>

  (in yago-wd-full-types) and

  <http://schema.org/Country> <http://www.w3.org/2000/01/rdf-schema#subClassOf>

  <http://schema.org/AdministrativeArea>

  (in yago-wd-class) to

  Germany type Country

  Country subclass AdministrativeArea

  Then, using this simplified file, we want to implement an algorithm to check for two classes ‘a’ and ‘b’ whether ‘a subclassOf b’ holds, directly or indirectly (this means applying transitivity and reflexivity to the original relation) and to determine (using the previous step), for entity ‘c’ and class ‘b’, whether ‘a type c’ holds (for instance, above we conclude that

  ‘Germany type AdministrativeArea’).

  In addition, we want to relate this to the information in the rest of YAGO, where type is sometimes implicit, as in

  <1st_Air_Corps_(Germany)>       <foundingDate> <1939>

  by finding all the cases where the type is included in the tag. Finally, we want to check how the hierarchy deals with events and actions; for instance, the RDF tuple

  <1887_U.S._National_Championships_(tennis)> <subEvent>

  <1887_U.S._National_Championships_–_Doubles>

  is included in YAGO, but not the fact that the 1887 U.S. National Championship is a type of ‘tennis tournament’.

  References Cited

  • “YAGO2: A spatially and temporally enhanced knowledge base from Wikipedia”, Johannes Hoffart and Fabian Suchanek and Klaus Berberich and Gerhard Weikum, Artificial Intelligence, v. 194, pages 28-61, 2013.
  • “YAGO 4: A Reason-able Knowledge Base”, Thomas Pellissier Tanon, Gerhard Weikum, Fabian M. Suchanek, Resource paper at the Extended Semantic Web Conference (ESWC), 2020.
  • Foundations of Semantic Web Technologies, Hitzler, Pascal and Krötzsch, Markus and Rudolph, Sebastian, CRCPress, 2009.

  SQL ON THE COMMAND LINE

  The Unix/Linux command line has many single-purpose operators that can be combined using pipe or input/output redirection to implement complex operations. A student of mine was able to implement basic SQL queries (SELECT … FROM … WHERE … statements) to run directly on files by translating the query to a pipeline of command line operations. However, the student did not have the time to test the command thoroughly or evaluate its performance. The purpose of this project is to do exactly this. The implementation is in Python. A second, more ambitious part is to prepare the program so that it can be incorporated into any Linux distribution that wants to do so.

  References Cited

  • The Linux Command Line: A Complete Introduction, William Shotts, No Starch Press, 2020.
  • Learning SQL: Generate, Manipulate, and Retrieve Data, Alan Beaulieu, O’Reilly, 2020.
  • Data Science at the Command Line: Facing the Future with Time-Tested Tools, Jeroen Janssens, O’Reilly, 2014.

  DATA ANALYSIS ON THE COMMAND LINE

  The Unix/Linux command line is a great tool for analyzing files; however, it has a few shortcomings that make it unable to tell us much about the data inside those files. One of them is the fact that, in a text file, numbers are treated as strings, so no arithmetic is possible. Hence, we cannot ask for the average of a column of numbers (we can simulate this with some effort). In R, a data file can be loaded and the system automatically recognizes numbers; the ‘summary’ command produces an analysis of the data according to its type. A student of mine developed a ‘summary’ command for the Unix/Linux command line. However, the command needs further testing and refinement. The goal is to have a command that will take a best guess as the type of data in a text file, classifying different variables (attributes, features) as numerical, dates, factors (categorical), and provide a summary for each. The command should be as efficient as possible and deal with files of all sizes.

  References Cited

  • The Linux Command Line: A Complete Introduction, William Shotts, No Starch Press, 2020.
  • Data Science at the Command Line: Facing the Future with Time-Tested Tools, Jeroen Janssens, O’Reilly, 2014.
  • R for Data Science: Import, Tidy, Transform, Visualize, and Model Data, Hadley Wickham and Garrett Grolemund, O’Reilly, 2017.

• Dr. Sabur Baidya

  

  Faculty Profile

  View Dr. Sabur Baidya’s faculty profile

  Research Areas:

  • Machine Learning
  • Edge/Fog Computing
  • Computer Systems

  Sample Project Description:

  Compressed Neural Networks for Multiscale Computing in Resource-constrained Systems

  Modern intelligent and autonomous systems are equipped with many sensors which are processed on single-board embedded computing platform. These systems often are battery powered with finite energy budget and they communicate with other devices for data/computation sharing over wireless networks. The constraint in computing, communication and energy, has led to the development of multiscale computing where data or computation can be opportunistically offloaded to an edge/fog server or to a cloud server [1]. However, if communication resources are insufficient then offloading data or commutation in raw form might not work and it necessitates compression of the data and/or computing algorithm. Additionally, with the boom of the Internet-of-things (IoT), devices are becoming lightweight which led to the development of TinyML [2] framework to support machine learning algorithms in small embedded boards and even on microcontrollers. However, depending on the application needs, the resource availability in the system and the constraints of available machine learning framework, the deep learning algorithms need to be compressed and/or optimally compiled to optimize the quality-of-service (QoS) of the applications.

  In this REU project, the undergraduate student will work on several Deep Neural Networks based applications for various sensor processing and investigate the following:

  • How can the models be compressed with some of the state-of the-art techniques, e.g., quantization [3], pruning [4], distillation [5], early-exit [6]. This will need studying the research papers to understand the theory behind the techniques and also implantation of the available open-source code for those techniques with Tensorflow or Pytorch on general purpose computer. This will need some knowledge of Machine learning, Python coding skills, and some experience of training ML models with Tensorflow or Pytorch.
  • Optimal compilation of the compressed models on different embedded platforms, e.g., Microcontrollers, Raspberry Pi, Nvidia Jetson Nano etc, with supported ML frameworks e.g., TVM, TF Lite, TF Lite Micro etc. The task will involve converting the ML flow graph to the supported ML framework for the corresponding platform. This does not need deep embedded system knowledge. Some experience on working with Unix/Linux based system will be good.
  • Create Pareto frontier with the accuracy, latency and energy performance tradeoff for various regime of computing resource availability on various platforms. This task will need measuring latency and accuracy of the algorithm, and energy consumption of the system.
  • Depending on availability of time, the student may try investigating the aforementioned characterizations for a few applications, e.g., Object Detection from image data, 3D object detection based on Lidar data etc.

  References Cited:

  1. Yousefpour, A., Fung, C., Nguyen, T., Kadiyala, K., Jalali, F., Niakanlahiji, A., Kong, J. and Jue, J.P., 2019. All one needs to know about fog computing and related edge computing paradigms: A complete survey. Journal of Systems Architecture, 98, pp.289-330.
  2. Warden, P. and Situnayake, D., 2019. Tinyml: Machine learning with tensorflow lite on arduino and ultra-low-power microcontrollers. O’Reilly Media.
  3. Hubara, I., Courbariaux, M., Soudry, D., El-Yaniv, R. and Bengio, Y., 2017. Quantized neural networks: Training neural networks with low precision weights and activations. The Journal of Machine Learning Research, 18(1), pp.6869-6898.
  4. Han, S., Pool, J., Tran, J. and Dally, W.J., 2015. Learning both weights and connections for efficient neural networks. arXiv preprint arXiv:1506.02626.
  5. Matsubara, Yoshitomo, Sabur Baidya, Davide Callegaro, Marco Levorato, and Sameer Singh. “Distilled split deep neural networks for edge-assisted real-time systems.” In Proceedings of the 2019 Workshop on Hot Topics in Video Analytics and Intelligent Edges, pp. 21-26. 2019.
  6. Teerapittayanon, S., McDanel, B. and Kung, H.T., 2016, December. Branchynet: Fast inference via early exiting from deep neural networks. In 2016 23rd International Conference on Pattern Recognition (ICPR)(pp. 2464-2469). IEEE.

• Dr. Dar-Jen Chang

  

  Faculty Profile

  View Dr. Dar-Jen Chang’s faculty profile

  Research Areas:

  • Compilers
  • Parallel processing
  • Graph analytics

  Sample Project Description:

  Using Python and its Libraries (e.g., PyTorch) for Big Data and High-Performance Computing Applications

  There are two data sets that can be used for such projects as described below:

  1. CelebFaces Attributes Dataset:

  • https://mmlab.ie.cuhk.edu.hk/projects/CelebA.html

  It is a large-scale face attributes dataset with more than 200K celebrity images, each with 40 attribute annotations. The images in this dataset cover large pose variations and background clutter. CelebA has large diversities, large quantities, and rich annotations, including:

  • 10,177 number of identities,
  • 202,599 number of face images, and
  • 5 landmark locations, 40 binary attributes annotations per image.

  The dataset can be employed as the training and test sets for the following computer vision tasks: face attribute recognition, face recognition, face detection, landmark (or facial part) localization, and face editing & synthesis.

  1. Lung Nodule Analysis 2016 data set:

  • https://luna16.grand-challenge.org/data/

  The data set includes 888 lung CT images and Annotation and Candidates csv files about the nodules present in the CT images. Sample projects of using the data set include:

  • Volume rendering
  • Image segmentation
  • Machine learning for lung cancer detection, etc.

  Graph Databases and Graph Analytics

  Based on graph theory and algorithms, graph databases and graph analytics are becoming popular and effective tools to analyze connected data and link them to machine learning algorithms [1]. A typical graph database and graph analytics project is given a dataset and study its applications by following this workflow:

  1. Dataset semantics study and preparation
  2. Graph database model design
  3. Import the dataset to a graph database such as Neo4j
  4. Basic statistics and graph exploration
  5. Applying graph algorithms like shortest path and centrality measurement
  6. Graph visualization
  7. Domain-specific applications like text sentiment analysis and recommendation

  There are two specific datasets identified here for this REU project, which are briefly described in the following sections.

  1. MIMC-III dataset ([2], [3])

  MIMIC-III (‘Medical Information Mart for Intensive Care’) is a large, single-center database comprising information relating to patients admitted to critical care units at a large tertiary care hospital” ([2]). This dataset is public accessible and can be used to study health care related issues.

  1. Yelp Open Dataset

  The Yelp dataset is a subset of our businesses, reviews, and user data for use in personal, educational, and academic purposes. Available as JSON files, use it to teach students about databases, to learn NLP, or for sample production data while you learn how to make mobile apps” (https://www.yelp.com/dataset). For example, [4] provides a study of Yelp 2018 open dataset using graph database and graph analytics.

  1. Covid-19 Dataset

  The complete Covid-19 dataset is a collection of the Covid-19 data maintained by Our World in Data (Our World in Data). It is updated daily and includes data on confirmed cases, deaths, hospitalizations, testing, and vaccinations as well as other variables of potential interest. The complete Covid-19 dataset can be downloaded in three file formats, CSV, XLSX, and JSON, from this link: covid-19-data/public/data at master · owid/covid-19-data · GitHub.

  References Cited

  • Mark Needham & Amy Hodler, Graph Algorithms: Practical Examples in Apache Spark and Neo4j, Published by O’Reilly Media, 2019.
  • Alistair E.W. Johnson, Tom J. Pollard, Lu Shen2, Li-wei H. Lehman, Mengling Feng, Mohammad Ghassemi, Benjamin Moody1, Peter Szolovits, Leo Anthony Celi1, & Roger G. Mark. MIMIC-III, a freely accessible critical care database, Sci Data 3, 160035 (2016).
  • Jessica A M Stothers & Andrew Nguyen. Can Neo4j Replace PostgreSQL in Healthcare? AMIA Jt Summits Transl Sci Proc. 2020;2020:646-653. Published 2020 May 30.
  • Kronmueller, M., Chang, D., Hu, H., and Desoky, A. A Graph Database of Yelp Dataset Challenge 2018 and Using Cypher for Basic Statistics and Graph Pattern Exploration, IEEE International Symposium on Signal Processing and Information Technology (ISSPIT), pp. 135-140, 2018.

• Dr. Anup Kumar

  

  Faculty Profile

  View Dr. Anup Kumar’s faculty profile

  Research Areas:

  • Mobile systems and Internet of Things (IoT).

  Sample Project Description:

  Cybersecurity Challenges in Internet of Things (IoT) Applications

  The internet of things has been a revolutionizing technology in the provision of healthcare, smart city, manufacturing and many other applications/services around the world. Enabled by a host of different technologies, the internet of things has spearheaded the development of many technologies for sustainable living, increased comfort and productivity for citizens as well as efficient operations for urban centers. IoT applications are enabled by collecting data through sensors within an application domain. The collected data is then processed followed by data mining to extract domain knowledge. However, there are several issues that pertain to how this data is sent and used. These issues include integrity, protection, and confidentiality of the data. In fact, this concern is not unwarranted as illustrated in the 2015 attack on the Ukranian power grid which left 225,000 people without power. Data gathered in IoT applications can be used to perform many undesired acts. This information may be used by bad actors to carry out unlawful acts causing risk to life and property. To secure the Internet of Things data and applications, typical security schemes might not be as effective and new approaches may need to be developed. In order to provide a standardized framework and terminology for discussing security attacks, attack incident taxonomy is suggested by the Computer Emergency Response Team (CERT) which was established by DARPA. There are different types of security and privacy issues in IoT applications. They exist on each of the three levels of the IoT architecture including application software layers, network layer and the perception layer. This research will map CERT attack taxonomy to IoT applications. In addition, the research provides holistic security solution approaches at all the three layers of IoT applications.

• Dr. Adrian Lauf

  

  Faculty Profile

  View Dr. Adrian Lauf’s faculty profile

  Research Areas:

  • Embedded systems
  • Accelerators

  Sample Project Description:

  Embedded Systems and Accelerators

  Embedded systems and microcontrollers are so ubiquitous in their application breadth that virtually all categories of devices and components seem to carry them. From disposable conveniences to durable goods and control systems, microcontrollers and embedded-class devices are everywhere. Thanks to significant efforts by microcontroller core designers such as Espressif, Atmel/Microchip, and ARM, anyone with a minimal amount of programming skill can implement basic system programming architectures with easy-to-use and friendly Integrated Development Environments (IDEs), some of which can even be run in a browser, short-circuiting traditional frustrations of importing libraries and dependencies [1]. However, not all IDEs and basic implementations are created equal, especially with respect to understanding power consumption and sleep states. As embedded devices are produced by the billions each year [2], even infinitesimally-minute differences can scale to produce enormous consequences when considering the collective carbon footprint of these devices. In respect to this, the student will:

  1. Write a set of programs that feature IO-based, timer/counter-based, and watchdog timer-based wakeup events for a microcontroller that runs a process control system. The process control inputs and outputs will be simulated by software and transmitted to the MCU via serial interfaces. Two control systems will be present – an industrial automation system (e.g., a sorting system or automated assembly plant), and a power generation station. The MCU implementation will represent one node in the process.
  2. These programs will be written in the top 3-4 prototyping/programming languages, including C, Wiring (Arduino), and MicroPython.
  3. Power consumption for each of the sleep state programs will be carefully monitored by directly cutting power feeds to the microcontroller and inserting a high-precision ammeter with a capture interface, provided to the student.
  4. The student will then extrapolate the power consumption differences between the language choices and IDEs to understand how these choices in aggregate can contribute to the global carbon footprint and power consumption based on the billions of devices produced annually.

  Additionally, these microcontroller architectures now feature fully-integrated communications modules, also easily integrated into an IDE through manufacturer-provided libraries. Examples include Espressif ESP 8266 and ESP32-based microcontroller implementations that feature WiFi, Bluetooth, BLE, and LoRa radios [3-5]. It is not uncommon for prototyping boards and even full-fledged devices to feature multiple radio types. It is also increasingly common that device-to-device mesh networks can provide some relief to infrastructure-based networks by distributing communications distances and loads between nearby devices [6]. With reference to carbon footprint and energy consumption, the student will be provided with an Espressif ESP32-based development board with Bluetooth, BLE, WiFi, and LoRa radios to perform the following tasks:

  1. A simulated process control network will be provided to the student complete with nodes intended for monitoring, control, and actuation. A set of 7 of these devices will be provided in various physical configuration spaces.
  2. The student will leverage the various communications interfaces based on data rate, power consumption, and security to create a multiple-radio network to increase reliability and range, and to reduce power consumption. Furthermore, sleep modes that enable/disable the radios as needed will be integrated to save additional power.
  3. Power consumption with the multi-radio implementation will be studied against a baseline of heterogeneous radio networks (e.g., all nodes use WiFi) to see what gains can be made given the network performance requirements.
  4. The information gathered from this set of experiments will be extrapolated to include process control networks around the world to understand the carbon and energy usage impact of embedded network choices in microcontroller applications.

  This work will assist in developing metrics and guidelines that can have a large cumulative impact by generating an energy consumption mapping of how microcontroller devices contribute to energy consumption based on configuration and radio usage.

  References Cited

  1. M. Banzi and M. Shiloh, Getting started with Arduino: the open source electronics prototyping platform. Maker Media, Inc., 2014.
  2. R. Lineback, “Microcontrollers Will Regain Growth After 2019 Slump ” IC Insights.
  3. Espressif SYstems, “ESP32 Datasheet,” Iot based microcontroller, 2017.
  4. U. S. Z. Abidin, “Design and development of multi-transceiver Lorafi board consisting LoRa and ESP8266-Wifi communication module,” 2018.
  5. C. Bouras, A. Gkamas, V. Kokkinos, and N. Papachristos, “Using LoRa technology for IoT monitoring systems,” in 2019 10th International Conference on Networks of the Future (NoF), 2019: IEEE, pp. 134-137.
  6. F. Adelantado, X. Vilajosana, P. Tuset-Peiro, B. Martinez, J. Melia-Segui, and T. Watteyne, “Understanding the limits of LoRaWAN,” IEEE Communications magazine, vol. 55, no. 9, pp. 34-40, 2017.

• Dr. Olfa Nasraoui

  

  Faculty Profile

  View Dr. Olfa Nasraoui’s faculty profile

  Research Areas:

  • Machine learning
  • Fairness in AI

  Sample Project Description:

  Explainable Machine Learning Algorithms

  The most accurate Big Data predictive methods tend to rely on black-box machine learning models, such as Deep Learning, which lack interpretability and do not provide a straightforward explanation for their outputs. Yet explanations can improve the transparency of a predictive system by justifying predictions, and this in turn can enhance the user’s trust in the system. Hence, one main challenge in designing a machine learning model is mitigating the trade-off between an explainable technique with moderate prediction accuracy and a more accurate technique with no explainable predictions.

  

  Explanation mechanisms play a critical role in building trust into human-machine learning algorithm interaction and can therefore contribute to building ‘fair’ machine learning algorithms.  Fair algorithms make predictions that are not biased against any group of people. In general most unfair models are the result of biased data. The rationale for ensuring fairness through explanations is that an unfair or biased prediction would be easier to detect when an explanation is provided along with the prediction.

  In this project, an REU student will investigate a set of white box and black box machine learning predictive models and methods to build in explanation ability and fairness in black box methods.  Experimental results should demonstrate that the developed method is effective in generating accurate and explainable predictions.

• Dr. Juw Won Park

  

  Faculty Profile

  View Dr. Juw Won Park’s faculty profile

  Research Areas:

  • Parallelization
  • Bioinformatics

  Sample Project Description:

  Large Scale Document Inversion using GPU Parallelization

  Current microprocessor architecture is moving towards multi-core/multi-threaded systems. This trend has led to a surge of interest in using multi-threaded computing devices, such as the Graphics Processing Unit (GPU), for general purpose computing. We can utilize the GPU in computation as a massive parallel coprocessor because the GPU consists of multiple cores. The GPU is also an affordable, attractive, and user-programmable commodity. Nowadays a lot of information has been flooded into the digital domain around the world. Huge volume of data, such as digital libraries, social networking services, e-commerce product data, and reviews, etc., is produced or collected every moment with dramatic growth in size. Although the inverted index is a useful data structure that can be used for full text searches or document retrieval, a large number of documents will require a tremendous amount of time to create the index. The performance of document inversion can be improved by multi-thread or multi-core GPU. Here I propose to implement a linear-time, hash-based, single program multiple data (SPMD), document inversion algorithm on the NVIDIA GPU/CUDA programming platform utilizing the huge computational power of the GPU, to develop high performance solutions for document indexing. The proposed parallel document inversion system will be implemented and evaluated using different test datasets from PubMed abstract and e-commerce product reviews. Junior level or senior level of undergraduate students with parallel programming experience will be a good fit for this project.

• Dr. Harry Zhang

  

  Faculty Profile

  View Dr. Harry Zhang’s faculty profile

  Research Areas:

  • Data-intensive computing
  • HPC
  • Data visualization

  Sample Project Description:

  Enabling Data-Intensive Computing on the Web with NSF XSEDE and NSF MRI

  As digital data sources grow in number and size, they pose an opportunity for computational investigation by means of data-intensive computing and analytics. Dr. Hui Zhang in NSF #1726532 has partnered with Texas Advanced Computing Center (TACC) to investigate IDOLS, a web application framework that allows users to customize the data-intensive workflows through simple configuration files as shown in Figure 1.

  

  

  To enhance the accessibility of large cyberinfrastructure to users from diverse domain fields, the IDOLS framework includes a set of pre-built task modules built at the University of Louisville and TACC to help bridge national researchers with remote hardware and software resources for data-intensive analytics and data-intensive computing. Utilizing this platform, REU students are expected to:

  1. To investigate a family of graphical user interfaces on IDOLS that can permit us to code, edit, and submit data-intensive analytics jobs empowered by a family of interactive web user interfaces;
  2. To migrate, verify and validate the IDOLS framework and learning modules at the University of Louisville’s in-house BDAP cluster, funded by NSF MRI project #1828521.

  Intellectual merit of the proposed project is embodied in a well-designed and implemented  cloud-based virtual laboratory (CVE) to introduce useful software tools for data-intensive computing and high-performance computing, and to lower the access barrier of state-of-the-art computing resources for students in STEM program and others whose research requires data-intensive analytics. Through this REU project, students will learn how to use data science tools and resources in their own projects. The summer research activity will give students a taste of the most compelling parts of a graduate school experience as well as an immersive experience in the field of high-performance computing and data-intensive computing. This research project will cover a breadth of skill sets in computing science. The students will gain experience in all aspects of the process of conducting research in a research lab. The summer project will help to address national interests by making state-of-the-art computing resources more accessible to students and teachers, supporting their development of critical workforce skills.

• Dr. Wei Zhang

  

  Faculty Profile

  View Dr. Wei Zhang’s faculty profile

  Research Areas:

  • Computer architecture
  • Compilers
  • Real-time systems

  Sample Project Description:

  Design Space Exploration for Computer Architecture Support for Deep Neural Networks

  Deep neural networks (DNNs) have been widely used for many AI applications including computer vision, speech recognition, self-driving cars, smartphones, drones, etc. While DNNs deliver high accuracy to enable AI in our daily lives, it comes at the cost of high computational complexity and intensive resource requirement [1]. These make it especially challenging to apply DNNs to edge and mobile devices, which have relatively limited hardware resources with stringent cost and/or battery constraints. While there are general-purpose and specialized computing architectures available for DNN such as TPU and GPU, they need to be customized to be used efficiently in edge and mobile domains for meeting the stringent performance, energy, and cost goals.

  In this project, the student will explore the architectural design space of computing platforms for supporting DNNs. A cycle-accurate architectural simulator [2] and/or analytic models will be used to evaluate the compute and memory bandwidth characteristics of different DNN applications and models. The student will evaluate the impact of different architectural designs such as on-chip network and memory hierarchy on performance, energy, and cost, and explore the most efficient computing architectures for running DNNs in a resource-constrained environment.

  References Cited

  1. V. Sze, Y. H. Chen, T. J. Wang, J. Emer. Efficient Processing of Deep Neural Networks: A Tutorial and Survey. Proceedings of the IEEE | Vol. 105, No. 12, December 2017.
  2. Francisco Muñoz-Martínez, José L. Abellán, Manuel E. Acacio, and Tushar Krishna. STONNE: Enabling Cycle-Level Microarchitectural Simulation for DNN Inference Accelerators, IEEE Computer Architecture Letters, Volume: 20, Issue: 2, July-Dec. 1, 2021.

• Eligibility Information

  To be eligible for the REU Site Summer Research Program for summer 2023, you must:

  • U.S. citizen, U.S. national, or permanent residents of the United States.
  • Majoring in Computer Science, Computer Engineering, or a closely related field, and has completed at least one year of study.
  • Minimum GPA of 3.0

  Women, underrepresented minorities, and students from institutions with limited research opportunities are especially encouraged to apply!

• Stipend, Housing, and Travel

  Stipend and Meal:

  • Participants will receive $7,000 stipend and meal allowance for this 10-week program.

  Housing:

  • Free on-campus housing (or housing allowance) will be provided.

  Travel:

  • $500 travel allowance to and from the UofL campus will be provided. Make sure you keep your receipts!

  Payment:

  • Students will receive their compensation in three equal payments: one at the beginning, one in the middle, and one at the end of the program.