Is Power Usage Effectiveness a metric only suitable for comparing a data-centre with itself?
Dr Ian F Bitterlin expounds.

The old adage about being able to improve only those things that you can measure has never been truer than in the data-centre arena.  As energy costs rise, carbon taxes become real, penalty charges loom and everybody wants to be seen to be ‘green’ the data-centre has (rightly, as a large consumer of power) attracted the attention of the sustainability band-wagon.  In a rush to be seen to be doing ‘something’ some trade organisations have produced metrics (measuring criteria) that can be applied to a data-centre to prove that it is not wasteful of energy.  If we ignore the sticky problem of ‘what’ the data-centre actually ‘does’ we are left with one metric that has found almost universal favour – Power Usage Effectiveness, or PUE.

Defining PUE

The definition of PUE could not be simpler, or more open to interpretation.
PUE is simply the ratio of the total power being supplied to the entire data-centre facility to the power consumed by the IT load.  So it is always greater than unity but the closer to unity it gets the better the data-centre is meant to be.  However it makes no judgement as to the ‘efficiency’ of the IT load itself.

The difference between the two power measurements includes such consumers as:

• Fans and pumps associated with the mechanical cooling system
• Compressor motors of chilled-water and other cooling plant
• Heaters and humidifiers
• Fans & cooling compressors associated with fresh-air make-up systems
• UPS efficiency losses
• Electrical distribution losses
• Security and Building Management Systems
• Internal & external lighting and cameras

However the losses are dominated by the first three on the list – i.e. the power consumption of the mechanical cooling plant.

One of the reasons why the definition of PUE is often blurred is best demonstrated by an example:  Take a small data-centre (incl NOC) consuming 800kW (900kVA @ 0.89 PF) located within a larger office complex that is separately consuming 720kW (900kVA @ 0.8PF).  The total load for the one facility is therefore 1520kW (1800kVA @ 0.84PF).  Now, inside the data-centre, the IT load consumes 445kW (445kVA @ Unity power factor).  If the facility manager can only measure the IT load and the incoming building supply the PUE appears to be 4.05 using the kVA data or 3.42 using the kW data.  If the data-centre load can be isolated from the rest of the building (often easier to say than to do) the PUE can now be calculated ‘properly’ – but we have two more possible answers, 2.02 by kVA and 1.80 by kW.  With results varying from 1.8 to 4.05 it is clear that confusion (and sometimes deliberate ‘spin’) is easy to create.

For completeness it is worth considering which elements are not included in the PUE metric:

• The efficiency of the micro-processor e.g. in terms of MIPS/Watt
• The efficiency of the software on the hardware platform e.g. in terms of Ws used/routine
• The efficiency of the IT loads’ on-board power supply (traditionally as low as 55% but now nearer 75% average, that is 25% losses)
• The power consumed by the IT loads’ cooling fans, often 5% of the total consumption
• Transmission and transformer losses in the grid
  

Mechanical strategies for PUE reduction

Click to enlarge

• Manage the air-flow
• Hot-aisle/Cold-aisle
• Deploy vented floor tiles in only the Cold-aisles
• Fit blanking plates
• Contain either the Hot or Cold-aisle
• Raise the server inlet temperature, e.g. from 20-22°C up to 27°C
• Raise the chilled water temperature, e.g. from 6/12°C to 12/18°C
• Remove all humidification (allowing 20-80%RH)
• Reduce (or remove) fresh-air make-up and associated conditioning to less than 0.1 air-change per 24h – treating the space as a machine room, not an occupied space
• Apply Variable Speed Drives to all fans and pumps
• Install free-cooling plant (avoiding the compressor cycle when the exterior ambient permits)
• Ultimately, apply fresh-air cooling with the air-conditioning plant turned ‘off’, in reserve
  

These strategies (all contained within the 2009 EU CoC &/or ASHREA TC9.9 2008 recommendations) will reduce the energy consumption in a data-centre from 5% to over 35%, reflected in PUE reductions from a typical 2.0 to as little as 1.3.

However the data centre operator may have commercial constraints on what can be done in the facility that prevent many of these energy saving strategies from being implemented.  For example a collocation facility may well have clients who will insist (based on fear or ignorance) upon a traditional SLA of 22°C±1°K and 50%±10%RH and that will be the end of any efficiency initiative.  At the other end of the scale is the huge homogeneous server load of a search engine provider that can go straight to the end game of fresh-air cooling and have no air-conditioning at all – just fans, filters and a high rate of external air-flow.

Electrical strategies for PUE reduction

There are electrical measures to be taken (some only applicable for our 3-wire 208V/60Hz North American cousins) that can also contribute significant energy savings (with the typical efficiency percentage improvements):

• 400VAC four-wire (3ph+N) power distribution negating the use of PDU transformers (2%) and using 230VAC 1Ph IT hardware instead of 120/127VAC (3%)
• 400VAC input/output UPS negating the use of output and bypass transformers (1.5%)
• Line Interactive transformer-less UPS (4-6%)
• Battery-free UPS negating the need for plant-room air-conditioning (2%)
• Naturally ventilated switchgear
  

Taking all these measures can reduce the power consumption by up to 15% - not as dramatic as that obtainable in the mechanical arena but nonetheless substantial.

So, what is the problem with using PUE as a ‘measure of goodness’ and comparison between data-centres?

PUE in the real world

The PUE can vary widely for a given facility based on four variables:

• Load
• Tier Classification
• Design & Operation
• Location

The design can incorporate all of the energy strategies listed above but can also be executed in a modular architecture – so better matching the load to the capacity and keeping the individual plant load high to gain valuable percentage points of efficiency.  In terms of operation it is a matter of house-keeping; making sure that all the energy strategies (such as blanking-plates etc) are followed to the letter.

As to the geographical location, the following points affect the potential PUE:

• Latitude & Altitude
• Ambient temperature, degree-days
• Without free-cooling
• With free-cooling
• City-centre
• Solar insolation (solar heat gain of the building fabric)

The reader is encouraged to consider that how facility that reports a PUE of 20 (just for example) can be a more efficient design, and therefore ‘better’, than a facility that reports a PUE of 2 at the same location .

This highlights one of the largest limitations of PUE as a metric and why it should never be used to compare facilities:  Energy efficiency, despite increasing importance, is not the most important feature of a data-centre design.  The designer must follow the business needs of the client and, in the very earliest days of the design process, must establish the degree of redundancy and concurrent-maintainability that the client needs/wants.  They may wish to use the Uptime/TIA Tier Classifications to benchmark the Availability level but, always, increasing the redundancy in plant and operational paths will increase the minimum possible PUE.

It should now be clear that it is not possible to compare the PUE:

• Of a Tier I/II facility with the PUE of a Tier III/IV facility at any load
• Of identical facilities in different latitudes (e.g. London and Edinburgh) &/or elevations (e.g. 100m and 1500mASL) at the same load
• Of a rural homogenous 50,000m_ server-farm and a city-centre 1,000m_ Colo-facility
  

So why do people, inside and outside of the industry, try to do just that?

The PUE metric is limited in its range but very useful in its application to a single facility.  It will, if monitored continuously (or at least quarterly) and published as an averaged annualised figure, serve to demonstrate the improvement in the effectiveness of the M&E services.  Any reported ‘improvement’ between adjacent quarters have to be ignored as the season will dictate the result not the facility.  It does, of course, let the IT hardware efficiency off-the-hook completely – a pity, since that represents the greatest opportunity to save energy.