1. The DataCenter is the Computer
Ron Hutchins, PhD
Associate Vice Provost for Research and
Technology, and Chief Technology Officer
Georgia Tech
2. Motivation and
Articulation of the Problem
• Simulation and modeling, as well as analytics, as
motivation today
• Growth of computing need for simulation is linear at
worst (best?)
• Hosting costs are growing, sustainability must be
considered
• A complex adaptive systems problem, not a simple one
dimensional issue: power/smart grid, cooling, outside
air, local water consumption/gray water, data center
network, monitoring, world-wide distribution of
partnered systems, chip clocking, black box vs manned,
software characteristics
3. The meta computer
• Large scale CPU consolidation into high density racking with in-row
cooling.
• Large storage arrays with multiple speeds/sizes separate from
Compute section including high speed scratch storage (like main
memory?)
• High performance networking switching can be compared to
backplane
• High speed connections with outside world
• Interactions across multiple data centers.
• Data center provides “heat sync” capabilities at large scale.
So… like we optimized the CPU chip by making it multi-core, L2 cache,
variable clock speed, etc., we need to work the same way with the
data center optimization – small pieces that can add up to a lot
5. cooling, outside air, local water
consumption/gray water
• Cold isle temperature recommendations are
changing. Air side economizing is becoming
more viable for more areas – requires careful
monitoring?
• Water is a big issue in the future. Gray water
capture/use is complex.
http://www.internap.com/colocation-
provider-facility-overview/green-data-centers/
7. chip clocking
“Designers now face a difficult choice between
increasing clock frequency to improve
performance and paying a large penalty in power
consumption, or reducing power with little gain
in the performance per gate of the design and
using more gates (silicon) for performance gains.”
http://www.eetimes.com/design/eda-
design/4018851/Greening-processor-design
8. data center network, monitoring
• Data center networks appear like WAN
networks due to the number of devices
attached – and the complexity of
interconnection.
• MPLS and VLANs rule the networks and create
complex architectures. Complexity is king.
• Collected data is basically for billing, not
operational/optimization purposes.
9. software characteristics
• Long running codes – sensitive to failures
• Highly parallel codes – sensitive to
interconnect
• Decoupled serial jobs – highly mobile
• Web services – stateless
• Database – hard to distribute and hard to
replicate in real time.
10. A Research Instrument –
A Systems Approach
• A production capable facility containing the best current capabilities
across multiple disciplines, a capability of being easily adaptable for
future innovations – including waste heat reuse in adjoining office
tower.
• Careful placement of sensors (air pressure, temperature, humidity,
power use) throughout the data center and in high performance
computer racks and chassis.
• Direct coupling of output of sensors to:
– Outside air controls
– Chip clocking controls
– Software schedulers
– Geographic prioritization
– Smart Grid/spot market for power
– Data center networking managed for optimization not billing