Open

Tinghui Zhang is a PhD student at the University of Florida, starting Fall 2024, with research interests focused on advancing language models, multimodal reasoning, natural language processing, knowledge graphs, and databases. He got his Master's degree in Computer Science from the University of Florida in 2023 with a GPA of 3.56, and Bachelor’s degree From FIU in 2020. His recent projects include developing algorithms that compress prompts and create scene graphs for video analysis. Before this, Tinghui has gained many different skills as a software engineer, working on different parts of both backend and frontend development in different projects. His technical skills include many programming languages and frameworks, which prepared him to create innovative solutions in his PhD research. He plans to use his extensive skills in his research and make significant contributions to the field of computer science.

Poster Abstract

Internal node cache size is critical to the ROXIE system. In the poster, we introduce advanced machine learning techniques to optimize the cache size settings, aiming to enhance the system performance.

The methodologies discussed include Bayesian Optimization (BO), Deep Learning (DL), and Reinforcement Learning (RL), which are evaluated for their ability to predict optimal cache configurations under varying system conditions. This analysis provides a detailed overview of state-of-the-art automated database tuning techniques. The proposal emphasizes the need for a simulated environment to gather data on cache performance for ROXIE, as actual performance data is unavailable. This simulation is designed to accurately reflect ROXIE's caching mechanisms and tests various parameters such as cache sizes and disk speeds to determine optimal settings for cache management.

Furthermore, the proposal outlines the development of a deep learning model trained with this simulated data to predict the best cache sizes for various configurations and workloads. This model aims to minimize the trial-and-error aspect of cache size tuning by providing more accurate predictions based on the simulated training data. The poster also discusses potential future expansions of this research by introducing techniques from other areas of database management and optimizations into the ROXIE system, reinforcing the application of machine learning techniques to the HPCC system for performance improvement.

Presentation

In this Video Recording Tinghui provides a tour and explanation of his poster content.

Learned Cache Size Setting for ROXIE:

Click on the poster for a larger image.

Open