DSpace Coleção:https://repositorio.ufms.br/handle/123456789/105402026-04-02T09:48:31Z2026-04-02T09:48:31ZAdaptive Load Balancing for Distributed LLM Inference Using Ollamahttps://repositorio.ufms.br/handle/123456789/138162025-12-05T11:44:44Z2025-01-01T00:00:00ZTítulo: Adaptive Load Balancing for Distributed LLM Inference Using Ollama
Abstract: Large Language Models (LLMs) have become crucial components of modern AI systems, supporting a wide range of applications through natural language generation tasks. While open-source serving platforms such as Ollama simplify the deployment of LLMs in local and on-premises environments, efficiently distributing inference workloads across multiple Ollama instances remains an open challenge. Conventional load balancing strategies, initially designed for stateless web services, fail to account for the dynamic execution characteristics of LLM inference, leading to suboptimal resource utilization and increased latency, particularly in heterogeneous computing environments. This paper proposes OllamaRouter, a specialized load balancing strategy tailored for distributed LLM inference using the Ollama text generation API. The proposed strategy dynamically allocates requests based on runtime estimates of token processing time and queue load, adapting to heterogeneous node performance without introducing significant computational overhead. To evaluate the effectiveness of OllamaRouter, we conducted a controlled experiment comparing its performance agains conventional Round Robin and Least Connection strategies, under a constant incoming request rate. The results show that OllamaRouter delivers higher throughput and lower average request latency, particularly as the number of waiting requests increases. The observed practical improvements and increased stability highlight the potential of adaptive load balancing in optimizing distributed LLM serving scenarios.
Tipo: Trabalho de Conclusão de Curso2025-01-01T00:00:00ZIntegração dos Coeficientes Cepstrais de Frequência Mel no Framework Mir_Refhttps://repositorio.ufms.br/handle/123456789/137072025-12-03T20:22:58Z2025-01-01T00:00:00ZTítulo: Integração dos Coeficientes Cepstrais de Frequência Mel no Framework Mir_Ref
Abstract: With the growing consumption of multimedia content, the field of Music Information Retrieval (MIR) has established itself as a strategic area for the analysis and interpretation of audio signals. In this context, the extraction of audio features, such as timbre, plays a central role in several tasks, but significant challenges remain due to the complexity and diversity of musical signals. Despite advances in the field, there is still a gap regarding the systematic comparison between traditional feature extraction methods and modern approaches based on deep learning. In particular, the challenge persists of integrating and comparing different types of audio representations within a standardized and reproducible experimental environment. This study aims to investigate the feasibility of integrating Mel-frequency cepstral coefficients (MFCCs) into the mir_ref framework, enabling the inclusion of this classical representation in comparative experiments with modern methods, such as embeddings generated by deep neural networks. In addition to the inclusion of MFCCs, the quality of the representations extracted by different methods was evaluated within the same experimental framework, promoting reproducibility and standardization in MIR evaluation. The conducted experiments demonstrate the impact of representation choice on music classification tasks, highlighting the particularities and limitations of each approach. Therefore, this study contributes to a comparative understanding of traditional and modern techniques for audio representation.
Tipo: Trabalho de Conclusão de Curso2025-01-01T00:00:00ZUma revisão de escopo sobre importância da neutralidade de viés e transparência dos dados em sistemas de recomendaçõeshttps://repositorio.ufms.br/handle/123456789/105412024-12-07T20:27:40Z2024-01-01T00:00:00ZTítulo: Uma revisão de escopo sobre importância da neutralidade de viés e transparência dos dados em sistemas de recomendações
Abstract: Recommender systems play a crucial role in personalizing experiences on digital platforms, influencing activities ranging from online shopping to content consumption. Despite their benefits, such as efficiency and convenience, these systems face criticism for biases, lack of transparency, and negative social impacts, raising ethical concerns regarding fairness and reliability. We conducted a scoping review to analyze how these issues have been addressed in the literature, with a focus on bias neutrality and data transparency. This review identifies 14 relevant studies and highlights that the primary types of bias include popularity, selection, and homogenization, while the lack of transparency undermines user trust and limits recommendation diversity. The findings indicate that mitigating these issues requires strategies such as diversifying recommendations, providing clear explanations to users, and implementing techniques to regulate biases. Furthermore, they emphasize challenges such as algorithmic opacity and the need for adaptable regulation. We conclude that developing ethical recommender systems requires integrating neutrality, transparency, and social responsibility, paving the way for algorithmic improvements and more inclusive policies.
Tipo: Trabalho de Conclusão de Curso2024-01-01T00:00:00Z