Cost Saving Strategies for Interactive AI Development
Lightningat Cloud Native + Kubernetes AI Day NA 2024 · on Tuesday, 2024-11-12· at 17:15· Salt Palace | Level 1 | Grand Ballroom A·
summary
Uses Apache YuniKorn to enforce fair sharing, user-specific policies and budget limits while carving an interactive compute pool (eg, 100 cores reserved when users are online) that returns resources when users go idle, plus GPU preemption and sharing to reduce wasted GPU time during exploration. Integrates with JupyterLab real-time displays so notebooks show live resource usage, and ties Kubeflow for training and Spark for data prep — delivering interactive guarantees without raising overall infrastructure cost.
AI-written summary · about
related talks
05mSponsored Keynote: The Cloud (And Your $$$) is Not Infinite: Dynamic Scheduling at Every Layer
31mAccelerate AI adoption: automated cloud efficiency
09mDemocratizing Access to GPUs for Faster MLAI Innovation with Generative AI
21mBest Practices for Optimizing Your Cloud Spend with AI
31mAI, CERN, and the Quest for GPU Custody: How CERN Leverages DRA for Efficient GPU Sharing
13mDesigning an Autonomous Workbench for Data Science on AWS
33mUncover Kubernetes cost-optimization trends
12mHow To Run AI At Scale Without Going Broke