ECE Master's Projects
Permanent URI for this collection
Browse
Recent Submissions
Item Dynamic task scheduling to improve performance and distribute heat uniformly on multicore system(Wichita State University, 2024-05) Pandi, Koteswara Rao; Asaduzzaman, AbuFrequent core utilization to achieve the required computation and communication goals results in high heat generation on the processor chip. Excessive heat generation could potentially cause the chip to overheat and malfunction. Multicore architecture with wireless network-on-chip (WNoC) has become a promising solution to satisfy the growing performance-power requirements. However, multicore WNoC with uniform or non-uniform subnets may suffer with poor core utilization, which reduces the performance and life span of the processor chip. In this work, we propose a dynamic task scheduling algorithm to improve performance and distribute heat uniformly on multicore WNoC architecture. The proposed algorithm consists of a two-step sorting process employing short-job-first (SJF) scheduling for uniform time distribution to achieve even heat dissipation, while maximizing resource utilization. A 49-core (7x7 WNoC) architecture with 1-Hop (where a core needs only one hop to reach the nearest router), 2-Hop, and 3-Hop architecture are modeled and simulated. The simulation programs are run using a representative synthetic workload, with 92 jobs, to evaluate the proposed algorithm. The simulation results exhibit that the 1-Hop architecture helps reduce the hop count by up to 45.83%, latency by up to 19.28%, and power consumption by up to 37.16%. The proposed scheduling algorithm helps distribute heat uniformly on the WNoC chip. Incorporating ‘the propose dynamic task scheduling algorithm for WNoC’ into graphical processing unit (GPU) and similar architectures could potentially create a formidable computing powerhouse, equipped to address an even wider range of computational tasks.Item Machine learning based car data prediction for network simulation using PETSC library(Wichita State University, 2023-12) Bogireddy, Sivaprasad Reddy; Asaduzzaman, AbuIn city areas, the profit of a business entity (i.e., shop) depends on the vehicular traffic at the nearest intersections. Argonne National Laboratory (ANL) and UChicago Argonne have developed Portable Extensible Toolkit for Scientific Computation (PETSc) and Data Management Network (DMNetwork) libraries to conduct network simulation. However, PETSc and DMNetwork do not provide a way to obtain and/or predict the number of vehicles to a network node. In this work, we develop a methodology to obtain real-time traffic data and predict future car data of the network nodes for profit analysis through network simulation. Open-Source Computer Vision Library (OpenCV) with the Background Subtractor algorithm and OpenCV with pre-trained You Only Look Once (YOLO) Version 3 datasets for vehicle detection are used for collecting car data using live video streaming file. The experimental results indicate that both methods yielded highly accurate traffic data, achieving a 90% accuracy rate, as cross-verified through manual counting for validation. We also integrate machine learning models, namely, Deep Neural Network (DNN) and Recurrent Neural Network (RNN), to forecast future car data for effective decision making. The DNN and RNN models are evaluated using three diverse datasets: Dataset-1 (represents a typical day car traffic), Dataset-2 (represents a morning high car traffic, and Dataset-3 (represents an evening high car traffic). Simulation results show that for both DNN and RNN models, Dataset-1 outperforms Dataset-2 and Dataset-3, achieving an R-squared value of 0.8 (indicating a high accuracy). Future work includes developing parallel tools for real-time car data retrieval, parallelizing file reading into PETSc DMNetwork, and incorporating additional features such as holidays, festivals, weekdays/weekends, special occasions, and weather forecasts for more robust future predictions.