FURI | Spring 2024

Optimizing Disaggregated Memory for Large Machine Learning Applications

Data icon, disabled. Four grey bars arranged like a vertical bar chart.

Large machine learning (ML) models have intensive memory demands for model training and serving, while memory resources in data centers are often underutilized. Memory disaggregation decouples memory resources from servers into a shared memory pool, and has the potential to expand memory capacity and reduce memory wastage. This project develops a memory disaggregation based caching system optimized for large applications with a real prototype and a comprehensive experimental evaluation. It offers insights into the proper utilization of disaggregated memory for ML models in pursuit of optimized performance and memory utilization.

Student researcher

Lillian Elizabeth Seebold

Computer systems engineering

Hometown: Queen Creek, Arizona, United States

Graduation date: Spring 2026