Current
DESC: Type I: Towards Greener AI Computing: Designing and Managing Sustainable Heterogeneous Edge Data Centers
(09/2023 - 08/2026)
NSF
This project is designed to significantly enhance the sustainability of edge data centers specialized in AI computing. Utilizing a comprehensive, carbon-first vertical approach, the project will address multiple levels of optimization. We will develop robust hardware accelerators that are energy-efficient, and also implement system-level improvements aimed at better power usage and strategic workload distribution. Additionally, renewable energy sources will be integrated into the management of these edge data centers. The ultimate goal is not only to reduce the overall carbon footprint through operational and embodied efficiencies but also to shift towards a more localized, renewable energy-based infrastructure. This holistic approach aims to pave the way for a new standard in energy-efficient, sustainable data center operations.
Collaborative Research: CyberTraining: Pilot: Research Workforce Development for Deep Learning Systems in Advanced GPU Cyberinfrastructure
(12/2022 - 11/2024)
NSF
With the recent advancements in artificial intelligence, deep learning (DL) systems and applications have become a driving force in multiple transdisciplinary domains. Meanwhile, this evolution has been supported by the rapid improvements of advanced GPU cyberinfrastructure (CI). Thus, DL system and application researchers need to constantly acquire interdisciplinary knowledge on advanced GPU CI. However, existing training materials are not only separately designed for a single aspect, either GPU CI or DL systems and applications, but also mainly focused on traditional techniques lagging behind quickly developed GPU CI and DL systems. To fill in this blank and overcome these difficulties, this project will develop a novel set of interactive training materials by involving six faculty members from five academic disciplines including computer science, computer engineering, data science, geospatial information science, and aerospace engineering.
Past
Enhancing synergistic execution of neural networks with joint pruning and mixed precision quantization
(01/2022 - 12/2022)
Center for Embedded Systems, NSF I/UCRC
Chip multiprocessors (CMPs) have become dominant in the automotive industry as they accommodate an increasing amount of cores in order to satisfy the increasing demands of the complicated services provided by modern vehicles. Additionally, heterogeneity has been introduced to combine performance and functional divergent offering significant computing power with more flexible power-performance trade-offs. As more and more services depend on concurrent execution of multiple neural networks, this multi-application execution environment on modern systems creates interference delays leading in violation of time constraints. In this project, we will extend the framework for synergistic execution developed last year, to further optimize the execution of neural networks by supporting joint pruning and mixed precision quantization.
Determinism of Critical Tasks on Multi-Core Platforms in the Presence of PHM Data Operations (project continuation)
(01/2022 - 12/2022)
Center for Embedded Systems, NSF I/UCRC
Prognostics and health management (PHM) data are collected on site but have to be eventually off-loaded to the cloud for further processing for predictive maintenance purposes. At the same time, the controllers of the system components should continue their operation unaffected by the PHM data collection and transmission. In this project, we plan to extend the framework built in the previous years in order to support more sophisticated functions and benchmark the effect of different services. Particularly, we will focus on the investigation of different FreeRTOS versions and compare their deterministic behavior against Linux with real-time kernel over a bare-metal hypervisor.
Enhancing vehicular applications using adaptive and synergistic resource management
(01/2021 - 12/2021)
Center for Embedded Systems, NSF I/UCRC
Chip multiprocessors (CMPs) have become dominant in the automotive industry. Additionally, heterogeneity has been introduced to combine performance and functional divergent offering significant computing power with more flexible power-performance trade-offs. As more and more services depend on concurrent execution of multiple neural networks, this multi-application execution environment on modern systems creates interference and delays leading in violation of time constraints. In this project, we will investigate and develop a service oriented architecture in order to allow synergistic resource management of neural networks on heterogeneous CMPs.
Heterogeneity-aware orchestration and efficient utilization of modern many-core systems
(01/2021 - 12/2021)
Center for Embedded Systems, NSF I/UCRC
When applications run concurrently on a CMP, they compete for shared resources, such as Last Level Cache (LLC) and main memory bandwidth. Conventional design approaches and industry trends are application agnostic. They mostly try to maximize consistency offering generic solutions and ignoring the requirements of diverse applications. In this project, we propose a two-step methodology to orchestrate the executing and resource allocation of modern CMPs. The goal of the first step is to estimate the performance of different applications on platform with different architecture, while avoiding exhaustive application profiling. The second step focuses on the automatic generation of performance estimators and resource allocation policies to be integrated into the scheduler so as to maximize the system utilization for specific workloads.
Determinism of Critical Tasks on Multi-Core Platforms in the Presence of PHM Data Operations
(01/2021 - 12/2021)
Center for Embedded Systems, NSF I/UCRC
Prognostics and health management (PHM) data are collected on site but have to be eventually off-loaded to the cloud for further processing for predictive maintenance purposes. At the same time, the controllers of the system components should continue their operation unaffected by the PHM data collection and transmission. In this proposal, which is based on the infrastructure that we have already built in the 2019 project, we investigate the effect of different multi-core configurations to provide PHM capability to an embedded system without impacting its performance. Additionally, we will explore the impact of data management services (e.g., encryption, compression, filtering) to the overall deployment in terms of deterministic behavior.
Real-time Deep Learning System on FPGA platforms for Autonomous Vehicle Applications
(01/2020 - 12/2020)
Center for Embedded Systems, NSF I/UCRC
This project investigates benefits and challenges on implementing deep-learning algorithms on state-of-the-art FPGA and GPU hardware platforms for real-time image classifications for autonomous vehicle applications. The two platforms will be compared in terms of toolchain complexity, inference accuracy, throughput, and power consumptions.
Modular and scalable infrastructure for the SIUC campus
(06/2019 - 06/2021)
Illinois Environmental Protection Agency through the US Department of Energy
Implement a scalable infrastructure consisting of photovoltaic (PV) panels and energy storage units to generate electricity in normal conditions and function as a backup power source in case of electricity outage. In the events of power outage, it will sustain the operation of a computing and control room in Engineering Building E as well as a wireless communication infrastructure. With this infrastructure and data obtained from its operation, the project intends to demonstrate that PV systems with energy storage provide a viable alternative to traditional diesel powered generators when selecting backup power sources for small-scale applications. In addition, the project will develop solar powered LTE communication modules to sustain cellular communication for emergency responders in the events of natural disasters that cause outages of both power and cellular service.
Prognostics and Health Management (PHM)-Enabled Real-Time Embedded Architectures
(08/2019 - 07/2020)
Center for Embedded Systems, NSF I/UCRC
Machinery (whether used in airplanes, automobiles, industrial machines, etc.) is inevitably subject to wear and needs to be maintained. Corrective maintenance, in which a component is replaced only after it has failed, is inapplicable/unacceptable in many cases where failures are catastrophic. Data-driven predictive maintenance is the smart way to do maintenance, but the collection of the PHM data should interfere the least with the controllers’ operation. The effect of different multicore architectures (core assignment, partition assignment) to provide PHM capability to an embedded system without impacting its performance need to be investigated.
Enhancing QoS of mixed criticality workloads on modern many-core systems
(08/2019 - 07/2020)
Center for Embedded Systems, NSF I/UCRC
The fact that cores share architectural components, such as caches and memory controllers, results in severe performance degradation. State-of-art approaches balance the application accesses to the shared resources by splitting equally the CPU time. Even though this policy is fair regarding the CPU time that each application gets, it does not take into consideration any contention effects that affect their performance. The focus of this project is to develop a four-step approach in order to (i) offer QoS guarantees by ensuring that high priority tasks and applications have minimal interference, and (ii) investigate the limitations of different Intel processor microarchitectures in order to offer such QoS guarantees.
Enhancing and Protecting Vehicular Applications Using Isolation and Adaptive Offloading
(08/2019 - 07/2020)
Center for Embedded Systems, NSF I/UCRC
The focus of this project is to develop techniques for enhancing and protecting vehicular applications using isolated application execution combined with transparent and adaptive task offloading. The first pillar is the application isolation in order to increase performance and meet time-requirements of automotive services. The second pillar is the trade-off estimation of task offloading as a way to further enhance quality of service, while the third pillar is the isolation of the applications increasing data integrity. The overall goal is to develop a unified design methodology and framework that will speed up and secure the content delivery and analysis process at the vehicle edge by using process isolation and task offloading.
Impact of Artificial Neural Network architectures on autonomous driving
(08/2018 - 07/2019)
Center for Embedded Systems, NSF I/UCRC
The rapid growth of on-vehicle multi-sensor inputs along with off-vehicle data streams provides an opportunity for developing innovative applications and services for modern vehicles. Autonomous cars employ hundreds of sensors for situational and environmental information and in order to integrate such services, new programming models and hardware infrastructure have been introduced. From the application perspective, machine learning algorithms have been utilized in order to further enhance the decision making process of autonomous vehicles. Image, speech and data classification are some examples where Artificial Neuron Networks (ANNs) prevail as the solution due to the complexity of the problem. In this project, we investigate the trade-offs of modern ANNs on embeded devices and the integration of neuromorphic platforms in smart cars.
Service oriented architecture and application optimizations for smart cars
(08/2017 - 07/2018)
Center for Embedded Systems, NSF I/UCRC
Modern cars employ hundreds of sensors for situational and environmental information and in order to integrate such services, new programming models and hardware infrastructure have been introduced. The project focuses on application optimization and development of service-oriented communication for seamless integration with regular automotive systems. Specifically, interfaces are being developed that will allow the communication of RTOS-based systems with more sophisticated software architectures and tools.
Environmental Information and Multi-Sensor Data Fusion Based Performance Estimations for Smart Cars
(08/2016 - 07/2017)
Center for Embedded Systems, NSF I/UCRC
The rapid growth of on-vehicle multi-sensor inputs along with off-vehicle data streams provides an opportunity for innovation in real-time decision making. The development of new advanced sensors is not sufficient enough without the utilization of enhanced signal processing techniques such as the data fusion methods. Multi-sensor data fusion (MSDF) is the process of combining or integrating measured or preprocessed data or information originating from different active or passive sensors. In this project, we develop and explore state-of-art data fusion techniques for decision making for automotive applications. Specifically we combine the benefits offered by car's increased connectivity (e.g. Internet, cloud services) with on-vehicle sensor information for providing detailed information about the state of the car and the environment.
Internet-of-Things Applications Development for Private LTE Small-Cell Networks (08/2016 - 07/2017)
Center for Embedded Systems, NSF I/UCRC
We are witnessing the dawn of a new era of Internet of Things (IoT; also known as Internet of Objects). Generally speaking, IoT refers to the networked interconnection of everyday objects, which are often equipped with ubiquitous intelligence. Smart building, self-driving cars, house monitoring and management, city electricity and pollution monitoring are some examples where dynamic networked real-time systems are deployed. This bloom of dynamic networked devices was also a result of new network services. With lower latency and higher bandwidth than its predecessor 3G networks, the latest cellular technology 4G LTE has been attracting many new users. However, the interactions among applications and network transport protocol still remain unexplored. In this project, we propose methodologies for interfacing, controlling and monitoring IoT devices in industrial settings over private LTE small-cell networks over 3.65 GHz frequency band.
Distributed Run-Time Management for a Multi-Agent System
(08/2015 - 07/2016)
Center for Embedded Systems, NSF I/UCRC
Today's prevalent solutions for modern multi-agent systems employ many processing inter-connected units leaving behind complex centralized approaches. Especially in modern automotive systems where high numbers of electronic components are employed. Navigation, car security, infotainment, travel information etc. are services highly depended on modern computing systems. In this project, we couple the concept of multi-agent systems with run-time resource management techniques in order to develop a distributed framework for run-time management of multi-agent systems.
Selected by NSF to be part of the “Industry-Nominated Technology Breakthroughs”.