SUMMARY

The challenge of efficiently managing the rapid growth of AI models and data, alongside volatile workloads, emphasizes the need for advanced storage solutions.

To address this, our research focuses on three main areas. First, we aim to design novel storage optimizations tailored for AI workloads, improving the handling of large and complex datasets. Second, we strive to identify optimal storage configurations by leveraging AI techniques, enhancing storage efficiency and security. Third, we focus on optimizing large language models (LLMs) within heterogeneous hierarchies to ensure efficient operation across complex and modern computing environments.

These research lines are crucial for advancing AI system capabilities amidst increasing demands and complexity.

OBJECTIVES

Within this domain, our current areas of interest include:

  • Storage Optimizations for AI Workloads: We aim to design and implement advanced storage mechanisms specifically tailored for AI workloads to handle large and complex datasets more effectively. By assessing bottlenecks in existing storage systems when handling AI workloads, we develop algorithms that optimize data placement, access patterns, and retrieval processes for AI applications.
  • AI-based Optimal Storage Configurations: While using AI-driven methods, we aim to find and maintain optimal storage configurations, ensuring efficient data management and security by offering self-tuning storage solutions. Our goal is to create adaptive storage systems that can dynamically reconfigure based on workload characteristics and performance metrics to enhance storage efficiency and data integrity.
  • Large Language Models (LLMs) Optimizations: We aim to enhance the performance and efficiency of LLMs operating in diverse and complex computing environments. By analyzing the interaction between LLMs, different levels of storage hierarchies, and CPU-GPU data and model transfer, we aim to develop techniques to optimize memory usage, data transfer, and processing speed for LLMs.

SELECTED PUBLICATIONS

  • Accelerating Deep Learning Training Through Transparent Storage Tiering.
    Dantas M, Leitão D, Cui P, Macedo R, Liu X, Xu W, Paulo J.
    IEEE/ACM International Symposium on Cluster, Cloud and Internet Computing (CCGrid), 2022.
  • The Case for Storage Optimization Decoupling in Deep Learning Frameworks.
    Macedo R, Correia C, Dantas M, Brito C, Xu W, Tanimura Y, Haga J, Paulo J.
    Workshop on Re-envisioning Extreme-Scale I/O for Emerging Hybrid HPC Workloads (REX-IO), colocated with Cluster, 2021.
  • Monarch: Hierarchical Storage Management for Deep Learning Frameworks.
    Dantas M, Leitão D, Correia C, Macedo R, Xu W, Paulo J.
    Workshop on Re-envisioning Extreme-Scale I/O for Emerging Hybrid HPC Workloads (REX-IO), colocated with Cluster, 2021.
Check the full list of publications here.

RESPONSIBLE FOR THE DOMAIN

Cláudia Brito

Assistant Researcher