AI4DL: Mining Behaviors of Deep Learning Workloads for Resource Management
summary
Presents a methodology for profiling and mining resource-usage patterns of containers running deep learning model training in IBM Deep Learning Services. Uses Conditional Restricted Boltzmann Machines (CRBMs) plus clustering to identify repeating phases and prototypical execution sequences, then applies phase-level statistical profiles to dynamically tune container CPU and memory allocations. Reports that this approach can cut CPU and memory use by about 30% versus reactive autoscaling while meeting resource demand over 95% of the time.
AI-written summary · about
chapters (10)
- 0:00Intro
- 0:24Context (Background & Motivation)
- 1:44Introduction
- 2:54Previous Work
- 4:23Methodology
- 5:13Phase Discovery and Detection
- 6:34Modeling Executions as Phase-series
- 8:06Resource Provisioning Policies
- 9:55Experiments: Phase Behavior
- 12:37Conclusions
related talks
19mResource Management Innovation for AI Workloads in Cloud Computing
35mAccelerate deep learning in the cloud with custom ML environments
19mML training acceleration with heterogeneous resources in Bytedance
40mOptimizing AI/ML Workloads
27mGreening the AI Cloud: Validating Power Models for Kubernetes Containers
32mLeveraging DRA for Conquering ML Workload Challenges on...
44mScaling ML from 0 to millions of users
12mElastic Resource Sharing for Distributed Deep Learning