Presentation by Mike Kelly CEO DataCentred at 2014 Data Centre World, asking does PUE miss the point?
Traditionally in data centres there has been no end to end view of energy efficiency. Chip designers have used techniques to make on-chip and on-board calculations efficient; rack builders have devised techniques which are efficient in cooling servers within racks and within containment systems for pods of racks; data centre operators have ensured that bulk efficiency of supposed compute energy demand versus total energy demand are as low as possible.
The reality, however, is that since no part of this chain has sight of the others, what is described as total efficiency, PUE, may bear no relation to compute efficiency at all. In large single user data centres, an end to end view is increasingly possible, and is a better measure of energy efficiency.
This presentation explores whether some of these end to end measures and some the efficiency techniques they lead to, could be used in a multi-tenancy data centre.
4. Does PUE miss the point?
• PUE measures energy consumption
• PUE concentration allows some energy saving
from historic baseline, ~30%, but real
compute efficiency could be very low
• Greater deficiency
– Assumes (in a sense) that all IT power
consumption is productive
– In reality actual use of silicon estate could be very
low, <10%
7. Single user data centre
• Single user will manage
data centre and compute
resource
• ‘Knows’ what’s happening
– can measure actual
compute efficiency
• Can measure compute
cycles per watt
• Can maximise efficiency
because can use shared
compute resource across
range of applications
11. Future Data Centres Internal
Environment – Radical OCP
• Form factors matched at
all scales
• Boards sit in racks as
moveable chassis
– Self supporting and
stackable
• Wireless networks
• Pluggable bulk connectors
for rack scale form factor
• ‘Racks’ stackable and
constructable for
reconfiguration and
convection friendly
shapes
• Robotic movement of
‘racks’ in high bay
warehouse like
environment
• Performance measures
IIOPs per watt, IPS or
FLOPS per watt
• Complete integration of
compute and internal
network with physical
operation of data centre
• Data centres as
neighbourhood hot spots
of pervasive compute
12. Future Data Centres in their
Environment
• Factories of continuous production
– Never idle
– Continuous renewal of equipment
– On site generation
• Neighbourhood resource like urban power
stations
• Integrated into CHP schemes
• Use of heat to create urban vegetation
production in biodomes in temperate zones
– Can achieve high radiative forcing levels
13. Compute Efficiency in Future Data
Centres
• Compute efficiency in data centres can’t be
achieved without use of local cloud
• Can’t be achieved without shared monitoring
• Can’t be achieved without shared resources
• Sharing of data centres is relatively novel
• Sharing compute systems is something IT
service providers ask their users to do all the
time
• Sauce for the goose
Editor's Notes
This is the future of data centres track so its about how the data centre can continue to develop to serve evolving customers needs
LINK – title of my presentation is “Energy Efficiency in Multi-Tenanted Data Centres” – and I want to explore how we can take the examples learned from the efficiencies of Google and initiatives such as the Open Compute project and apply to multi-tenant data centres.
First question is– so how do we measure energy efficiency?
Script from traditional approach slides in appendix
Script as per “The Cloud Begins With Coal – Big Data, Big Networks, Big Infrastructure, and Big Power", Digital Power Group report
Link -
- Indeed More efficiency in IT load could raise PUE
Processors and disks and networking cards consume very similar amounts of power whether they are doing something or sitting idle
So if your equipment is doing nothing then even if it is doing it efficiently then it is still doing nothing…
We’re proposing that we need to focus on compute efficiency as well as PUE
So how do we do this
INTRO – its relatively easy in a single user data centre
The big names- Google/ Facebook and the like can measure the full picture
Image is compute efficiency per watt graph
Can measure PUE, er
Could measure compute efficiency if tenants chose to ‘subscribe’ to a monitoring service (unlikely)
Arguably could price according to compute efficiency (hoho)
No real incentive for a data centre operator to be interested in compute efficiency
Local tethered cloud could
Standard colo plus availability of suitably structured and shared compute resource
Guaranteed to be used heavily
simple flexible resource
general architecture to suit tenants
Could be multiple architectures
Useful name, but refers to many different things for many different users
In multi-tenanted data centres sort of cloud
More like overflow/flex capacity
But guaranteed to improve efficiency
Bound to be diversity of demand between clients