Theses and Dissertations from DePaul University

Date of Award

Summer 2026

Degree Type

Thesis

Degree Name

Master of Science (MS)

Department

Computer Science

College

College of Computing and Digital Media

First Advisor

Alexandru Orhean

Abstract

For decades, advancements in information retrieval technologies have changed how individuals discover data with computer systems. More recent advancements in artificial intelligence (AI) have also made the benefits of neural information retrieval systems available to millions. These information retrieval systems help empower future discoveries, though this is not the case for scientific data and High-Performance Computing (HPC) systems. HPC systems can generate enormous amounts of data, and no tools are currently available that can ingest data at the rates required to efficiently build an information retrieval system to explore scientific data. This thesis explores the current state of constructing a neural information retrieval model from a systems perspective. Prominent tools, such as vLLM and FAISS, used to generate vector embeddings and to construct vector-based indexes are evaluated in depth. Even today, such tools struggle to generate the throughput required to meet HPC demands. In particular, the gap in required throughput is especially prevalent when generating vector embeddings utilizing state-ofthe-art Large Language Models (LLMs). To bridge the gap, this work proposes the design of a pipelining technique applied to Graphical Processing Unit (GPU) computations that can lead to improved vector embedding generation throughput and evaluates the potential performance gain through the implementation of a micro-benchmark.

Download

Share

COinS