Knowing the Energy Footprint
This can be measured in a variety of manners, from the macro (data centre, rack, cooling and a/c, or nodes for runs of a large HPC simulation) to the micro (elements such as FPU and data transfer between cores or over interconnect).
It is frequently useful to use an energy profiler to determine which contributions to a simulation consume most energy - is it the IO or perhaps a given subroutine/function?
Measuring the Energy Footprint
Leaving aside embodied emissions, there are several options
- read the meters (or energy bill) of a data centre
- use the in-built meters from a compute node
- access the performance counters on a modern microprocessor, using performance analysis tools (such as TAU, Allinea MAP or TAU)
- use of Intel's RAPL (for Xeon, and albeit limited for Xeon Phi (KNC, KNL) too)
- use of on-board sensors, read over IPMI (or perhaps SNMP) interface
- measure the voltage drop across components using kit such as MAGEEC. Typically this approach can be invasive.
Related Articles
- HEC's Greener Compute support pages and news & views
- Latest tweets on #greenerCompute
- Who's who from Adept to Oxford
- Some high level approaches to implementing Greener Compute
- Discussion on what to measure and ideas on how to do so
- Determining the Environmental Footprint of High End Compute
- UK's Greenest High End Compute (2014)