Electrical Efficiency Measurement for Data Centers

Overview

Using a fuel-economy analogy, the authors separate design specification, peer benchmarks, time-averaged efficiency and instantaneous efficiency, and define 'instantaneous' PUE as a one-hour average, since device cycling periods are tens of minutes. They illustrate the trap: data center A reads PUE 2.0 at 10% load while B reads 1.7 at 70% load, yet A has the better efficiency curve. A design specification example is expressed as PUE at 75% IT load and 25 deg C outdoor temperature.

For a typical dual-path data center with N+1 CRACs at 30% load the breakdown is IT 47%, chiller 23%, CRAC/CRAH 15%, UPS 6%, humidifier 3%, PDU 3%, lighting/aux 2% and switchgear/generator 1%, giving PUE 2.13. The paper also notes that a 1 MW high-availability data center can consume $20,000,000 of electricity over its lifetime.

The proposed approach is to calibrate a mathematical loss model of the facility from a limited set of meter points, then use ongoing measurements to detect deviations from the model. Using UPS output power as the IT load introduces a 2% to 25% error (PDU losses and other UPS-fed loads), which the model can subtract. Ongoing monitoring needs about half the number of measurement points used for the initial assessment, and periodic audits are ideally done twice a year, summer and winter.

Each one-time measurement should record date, time, averaging interval, meter points, rated and actual IT load, outdoor temperature and humidity, redundancy, heat rejection type, and economizer type and status. The paper also lists the contents of an annual efficiency report. It is a measurement method and set of principles, not a binding threshold for any one site.

Key points

1. A 1 MW high-availability data center can consume $20,000,000 of electricity over its lifetime.
2. Typical dual-path N+1 CRAC site at 30% load: IT 47%, chiller 23%, CRAC/CRAH 15%, UPS 6%, humidifier 3%, PDU 3%, lighting/aux 2%, switchgear/generator 1%; PUE 2.13.
3. Device cycling periods are tens of minutes, so the shortest meaningful PUE averaging period is about one hour.
4. A reads PUE 2.0 at 10% load, B reads 1.7 at 70% load, yet A has the better efficiency curve.
5. Using UPS output as IT load introduces a 2-25% error, which the model can subtract.
6. Ongoing monitoring needs about half the measurement points of the initial assessment; audits ideally twice a year (summer and winter).
7. Each measurement should record date, time, averaging interval, meter points, rated and actual IT load, outdoor temperature/humidity, redundancy, heat rejection type, economizer type and status.
8. A design specification example is PUE at 75% IT load and 25 deg C outdoor temperature.

Using it in design

• Never output a single PUE: produce a PUE-vs-IT-load curve (10-100%) and an annual PUE from hourly weather bins, labelling the averaging period explicitly.
• Store design PUE with its reference conditions (load %, outdoor temperature) and use the default loss breakdown to seed a legacy-baseline comparison.
• In the metering plan generator place permanent meters at the utility incomer, UPS outputs and major cooling feeders; compute IT load as UPS output minus modelled PDU loss and warn that raw UPS output overstates IT load by 2-25%.
• Add a commissioning/operations template with the 11 metadata fields and an annual report section listing actual versus expected PUE, IT load %, subsystem breakdown, kWh and cost; require permanent instrumentation in new builds.

Who should read it

Operations engineers, commissioning teams and energy managers who report PUE or set up metering plans.

Related standards

• ISO/IEC 30134-2
• EN 50600-4-2
• The Green Grid PUE

Publication details

Authors

Neil Rasmussen

Publisher

Schneider Electric

Year

2011

Document number

WP 154

Document language

English

Source type

Manufacturer

Access

Free to read

Licence

Copyright retained; citation only — the publisher’s terms apply to the original document

Citation