Date of Award
Fall 11-20-2024
Document Type
Thesis (Undergraduate)
Department
Computer Science
First Advisor
George Cybenko
Abstract
In the field of communications technology, efficient allocation of network resources is critical. Accurately predicting network traffic requirements such as packet demands is a significant challenge. Current core network utilization is typically quite low due to overprovisioning, as network operators only use a fraction of their capacity because they are unsure how to optimize traffic allocation while maintaining a high quality of service for customers. For example, in a 2021 survey of 500 IT professionals in the United States, 78% admitted that overprovisioning led to the discovery of further performance bottlenecks. The current commonly used equal-cost multipath (ECMP) routing strategy for managing multiple paths is also inefficient. This thesis proposes to address this challenge by leveraging the capabilities of Large Language Models (LLMs) to predict future network traffic. The main objective is to develop a model that can forecast the volume of packets that network users will require in the near future, enabling a more efficient and tailored allocation of network resources. The successful implementation of such innovative methods could significantly change how communications technology companies allocate network resources, leading to more reliable and user-centric network management. The results of this research demonstrate the potential of the lightweight nanoGPT model to address key challenges in network traffic prediction for resource allocation. The trained nanoGPT model demonstrates strong fidelity in its generated data, closely matching patterns in real-world network flows. Additionally, comparisons between generated and actual future flows using real-world data indicate that the model consistently captures complex relationships within the data and makes meaningful predictions. The evaluation of nanoGPT on multiple datasets obtained from networks at the Universidad Del Cauca, Popayán, Colombia in 2017 and 2019 demonstrate its robust predictive capabilities. The model even achieves promising results when tested on the 2017 network dataset with fewer applications (75 versus 141 on which it was trained). Further, nanoGPT achieved lower RMSE and MAE values compared to traditional models such as RNN and LSTM referenced in another study. Euclidean distances between actual and predicted data also show significant evidence of predictive power. These results showcase nanoGPT’s ability to generalize effectively, making it a promising tool for predictive tasks such as dynamic routing and congestion management in modern networks.
Recommended Citation
Krivov, Nicholas A., "INTELLIGENT MULTIPATH ROUTING: OPTIMIZING NETWORK RESOURCE ALLOCATION WITH LARGE LANGUAGE MODELS" (2024). Computer Science Senior Theses. 51.
https://digitalcommons.dartmouth.edu/cs_senior_theses/51
