In the fast-moving world of machine learning, efficiency is key. A major challenge is resource imbalance, with servers often having mismatched compute power and storage capacity. By using smart scaling based on real-time metrics, we can teach the compute framework to adjust their own resources, like GPU and CPU compute power and storage, just right.

Lu and Chunxu will show how the compute framework can be taught to self-adjust and become more efficient. This method promises not just speed, but also better use of resources and cost savings.

Topics of Discussion:
- Examining the issues of over or under-resourcing and how auto-scaling can fix it
- Exploring the development of self-adjusting compute and storage frameworks
- Training models to instruct the compute framework to use pod-level and application-level metrics to decide how to adjust its compute power and storage capacity
- Sharing actual success stories where auto-scaling has made learning faster and less expensive