AEGIS system optimizes deep learning training on shared GPUs

Researchers have developed AEGIS, a runtime scheduling system designed to improve the efficiency of deep learning training on shared multi-GPU servers. AEGIS manages the collocation of multiple deep learning workloads by integrating memory feasibility checks, post-placement observation, and runtime-pressure filtering. This approach aims to reduce resource underutilization and queueing times compared to exclusive allocation, while also mitigating performance degradation and out-of-memory failures that can occur with less sophisticated collocation methods. Evaluations using various workloads demonstrated that AEGIS can reduce training makespan by up to 27% compared to exclusive allocation and by 16-21% compared to other collocation systems like Lucid and Horus. AI

IMPACT Optimizes GPU utilization for deep learning training, potentially reducing costs and accelerating development cycles.

RANK_REASON The cluster describes a research paper detailing a new system for optimizing deep learning training infrastructure. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.LG →

• AEGIS
• arXiv
• Ehsan Yousefzadeh-Asl-Miandoab
• graphics processing unit
• Horus
• Lucid
• transformer

• infra
• paper

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →