Improving High Throughput Computing’s Energy Efficiency - a Measurement-Based Case Study
summary
Measurement study runs ATLAS HTCondor workloads in 16-core/32-thread Intel and AMD servers, reads power via SNMP from two PDUs, and times jobs across DVFS and core-count settings. Key findings: baseline idle power accounts for >80% of per‑job energy, active energy is nearly constant so raising frequency or core count shortens runtime and often lowers total energy; AMD jobs cluster around ~28 Wh per job. Recommendations: focus on cutting baseline draw (unplug unused devices, tune firmware) and use carbon-aware scheduling rather than wholesale server swaps.
AI-written summary · about
chapters (8)
- 0:00Introduction and context
- 2:12Project background and goals
- 8:42Power and energy modeling
- 11:34Experimental setup and results
- 17:20Key findings and trade-offs
- 20:30Recommendations and policy suggestions
- 23:05Conclusion and future work
- 24:11Q&A session
related talks
15mOptimizing Application Performance on Kubernetes
07mKeynote: Finding Your Power to Accelerate to a Sustainable Future
49mThe Future is 💚: A Journey through the Energy Efficient ..
21mHPC Green Computing: Relationships Between Throughput and Latency
18mPower Savings & NFV: How to reconcile these 2 constraints ?
20mMonitoring Energy Consumption and Building Energy Efficient Systems the Cloud Native Way
47mEnergy Proportional Edge Computing Infrastructure
46mNew Innovations in Power and Performance Management