New Innovations in Power and Performance Management
summary
Overview of concrete engineering: POWER systems expose ~40 sensor types (≈400 sensors) with OCC reading power every 500 μs and temperatures every 4 ms, pushing key values every 10–100 ms; a new OpenBMC out-of-band websocket push (crasSD) streams changes to a central service and scales to ~300 nodes at ORNL on AC922 hardware. A workload optimized frequency (WOLF) controller clips Linux frequency votes based on power/thermal headroom, and GEOPM node-level agents run hierarchical optimization to rebalance CPU/GPU/memory power.
AI-written summary · about
chapters (8)
- 0:00Start
- 2:36Hardware Block Diagram
- 4:11Workload Optimized Frequency (WOF)
- 9:23OCC Voting Box-P9
- 14:14Measurement-In Band Sensors IBM.
- 16:45Cluster RAS Service Daemon (crassd) IBM.
- 33:16Motivation: Performance Variability
- 40:58HPC PowerStack High Level Flow
related talks
43mOptimizing Full-Stack Sustainability in a Real World Data...
26mIndustrial Machine Performance Insights with AWS IoT Services
42meBPF-Powered Observab...
35mSustainability Research the Cloud Native Way
20mMonitoring Energy Consumption and Building Energy Efficient Systems the Cloud Native Way
47mEnergy Proportional Edge Computing Infrastructure
43mIndustrial Internet of Things Applied to Energy Efficiency
32mImproving High Throughput Computing’s Energy Efficiency - a Measurement-Based Case Study