Power Estimation of Scientific Workflow Tasks

Magnus Reid

L4

Motivation

• Power models are useful
  • Energy efficient scheduling (e.g. GreenHEFT)
  • Carbon aware capping
• Different tasks have different power characteristics (CPU Usage, I/O)
• Major dynamic range of power consumption is from CPU utilisation
• Scientific workflows have certain properties
• Don't always have extensive profiling from previous runs
• Don't always have external power monitoring equipment

Solution

• Power modelling of specific workflows and target infrastructure
  • Split workflow inputs
  • Model power from small sections of input
  • (Lotaru approach applied to power consumption)
  • Model effect of CPU throttling on task power consumption
• Use power model to limit power consumption
  • CPU throttling

Plan

• So far, have experimented with:
  • Power consumption measurement with RAPL
  • Workflow power consumption
    • Real world tasks
  • Effect of CPU throttling on power consumption
• Focus on building power model initially
• Investigate if using power model could help to prevent carbon emissions using CPU throttling