Authors:
Pedro Henrique Sachete Garcia
1
;
Ester de Souza Oribes
1
;
Ivan Mangini Lopes Junior
1
;
Braulio Marques de Souza
1
;
Angelo Nery Vieira Crestani
2
;
Arthur Lorenzon
3
;
Marcelo Luizelli
1
;
Paulo Silas Severo de Souza
1
and
Fábio Rossi
2
Affiliations:
1
Federal University of Pampa, Brazil
;
2
Federal Institute Farroupilha, Brazil
;
3
Federal University of Rio Grande do Sul, Brazil
Keyword(s):
Cloud-Native Applications, Digital Twins, Large Language Models, Resource Allocation.
Abstract:
Efficient resource allocation in programmable datacenters is a critical challenge due to the diverse and dynamic nature of workloads in cloud-native environments. Traditional methods often fall short in addressing the complexities of modern datacenters, such as inter-service dependencies, latency constraints, and optimal resource utilization. This paper introduces the Dynamic Intelligent Resource Allocation with Large Language Models and Digital Twins (DIRA-LDT) framework, a cutting-edge solution that combines real-time monitoring capabilities of Digital Twins with the predictive and reasoning strengths of Large Language Models (LLMs). DIRA-LDT systematically optimizes resource management by achieving high allocation accuracy, minimizing communication latency, and maximizing bandwidth utilization. By leveraging detailed real-time insights and intelligent decision-making, the framework ensures balanced resource distribution across the datacenter while meeting stringent performance req
uirements. Among the key results, DIRA-LDT achieves an allocation accuracy of 98.5%, an average latency reduction to 5.3 ms, and a bandwidth utilization of 82.4%, significantly outperforming heuristic-based, statistical, machine learning, and reinforcement learning approaches.
(More)