Special Studies on Power

1
Special Studies on Power Cooling Practices,
Requirements, Issues and Concerns for Data
Centers
Roger A Panton, Avetec Executive Director
DICE HPC User Forum Sept 2009 rpanton_at_avetec.org
2
Background The Growing Issue

• Power cooling demand is a world-wide major
  concern
• Exponential growth in HPC system size and
  performance have increased energy requirements
• Estimates are that the energy use has increased
  nearly 100 for data centers over the last six
  years
• If data center energy usage continues at the
  current rate, the nation will need to build two
  large power plants per year to meet demand
• Energy prices have risen substantially
• Aging data center infrastructure is an issue
• In this perfect storm the challenge is how to
  increase energy efficiency without compromising
  performance improvement

3
Two Related Studies on Power Cooling

• Study 1 Power Cooling Practices and Planning
  at HPC Data Centers
• Status Completed late Spring 2009
• REPORT AVAILABLE TO ANYONE WHO IS INTERESTED!
• WWW.DICEPROGRAM.ORG
• Study 2 Power Cooling Solutions for
  Enterprise and HPC Data Centers
• Status Complete report by end of September 2009
• Both studies are being conducted via a new
  Avetec-IDC partnership

4

• Phase 1 StudyPower Cooling Practices and
  Planning at
• HPC Data Centers

5
Background and Objective

• Avetec was tasked to evaluate power and cooling
  planning, constraints, requirements and how HPC
  data centers are coping
• Avetec contracted with IDC to survey HPC data
  centers and vendors to assess
• Current power and cooling situation
• Planning in place to address requirements
• Solutions currently being used
• Forecasted solutions over the next three to five
  years
• Respondents included HPC data centers and vendors
  including HPC systems, equipment and facilities

6
Survey Sample and Methodology

• Survey includes 41 respondents
• 28 HPC data centers
• 13 vendors of HPC products and services
• Response rate was approximately 40 of sites
  contacted
• Respondents were from the US, Europe and Asia
• HPC data centers were selected from the Top500
  list
• Centers selected fell between number 50 and 250
• Interviews were conducted by phone or in person
• Respondents also had the option to complete the
  survey on their own

7
Sample Respondent Mix
HPC Data Centers
HPC System and Service Vendors
8

• Study 1
• Highlights

9
Current HPC Data Center View

• Over 96 considered green design important
• Majority of sites expect power and cooling to
  impact future HPC center planning
• Majority of respondents have studied or
  implemented greener operations
• Most centers have used software models to analyze
  heat flow and/or power consumption
• Approximately half of the centers paid for power
  and cooling out of their budgets

10
General Findings

• HPC data centers average per site
• Available floor space over 26,000 ft2
• Used floor space about 17,000 ft2 (63 of
  available space)
• Cooling capacity 22.7 million BTUs or 1,839 tons
• Annual power consumption 6.356 MW
• HPC data center cost
• Annual power cost was 2.9 million or 456 per KW
• Ten sites provided the percentage of their budget
  spent on poweraverage was 23
• Two-thirds of the sites had budget for power and
  cooling upgrades
• Average amount is 6.87 million

11
Challenges and Expansion Constraints

• Majority of centers are starting to consider
  power and cooling efficiency equal to or more
  important than HPC computing performance
• Power and cooling issues are becoming the biggest
  barriers to expansion and upgrades
• Most HPC vendors are starting to see power and
  cooling as a brake on performance
• Power and cooling costs are becoming a key factor
  in upgrade decisions

12
Current Approaches

• Power and cooling are becoming a key factor in
  upgrade decisions
• Majority of centers have accomplished an air flow
  analysis to improve air cooling efficiency
• Majority of centers have analyzed power
  consumption
• Use of chilled water for cooling is increasing
• The power and cooling issues are being discussed
  across the HPC community
• Other approaches include
• Hot/cold aisle containment
• Re-directing/reusing excess heat
• Facility modifications/equipment refurbishment
• Raising temperatures

DRILLING DEEPER INTO THIS IN STUDY 2
13
Future Solutions and Technologies

• Approximately two-thirds of centers plan to
  expand or build new data centers
• About half of the data centers have or are
  planning to distribute HPC resources
• Liquid cooling is most frequently cited future
  alternative being considered (centers vendors)
• Other potential solutions included combination of
  air/water cooling
• In-row cooling
• Use outside air ambient

DRILLING DEEPER INTO THIS IN STUDY 2
14

• Specific Questions

15
Belief in Game Changing Technologies
Question Do you anticipate any changing cooling
technologies that will significantly change the
way HPC systems and data center are cooled in the
next 5 years?
Response Government Industry Academia All Sites
Yes 62.5 50.0 8.3 35.7
No 25.0 50.0 58.3 46.4
Not certain 12.5 0.0 33.3 17.9
Total 100 100 100 100
Government respondents more optimistic about game
changing technologies
16
Power and Cooling by Sub-systems

• Question How do your power and cooling costs
  divide among your HPC compute, storage, and
  visualization sub-systems?

Response Government Industry Academia All Sites
Compute 92.6 81.7 90.1 89.7
Storage 4.9 18.3 8.2 8.6
Visualization 2.1 0.0 1.1 1.3
Dont know or not sure 0.4 0.0 0.6 0.4
Storage costs greater among non-government sites
17
Expected Changes by Sub-systems

• Question In which of the three basic HPC data
  center sub-systems defined do you expect power
  and cooling to grow the most rapidly in the next
  five years?

Response Government Industry Academia All Sites
Compute 87.5 25.0 75.0 64.3
Storage 12.5 50.0 0.0 17.9
Visualization 0.0 12.5 0.0 3.6
Dont know or not sure 0.0 12.5 25.0 14.3
Total 100 100 100 100
Government academic centers forecast most
growth in compute whereas industry sees most
growth in storage
18
Guidance for Users

• Improving a site's power and cooling profile is a
  multi-dimensional process of stepwise refinement
  that begin at any time
• The most efficient and progressive HPC data
  centers surveyed are already using metrics to
  measure their performance in power and cooling
• A handful of power and cooling efficiency drivers
  will govern how HPC data centers take to greater
  efficiency
• HPC system upgrades will stimulate demand for
  detailed efficiency analysis services

19
Guidance for Vendors

• The recession will elevate the importance of the
  "green" design features of HPC systems,
  especially in Europe
• The next round of HPC system upgrades will
  stimulate demand for HPC data center efficiency
  analysis services
• Future multi-Pflop HPC data centers will require
  as much power as a small to mid-sized city and
  stimulate greater public policy concerns

20

• Public Policy
• Implications
• from Study 1

21
Public Policy Implications for Discussion

• Should the community take a proactive position
  through collaborative discussions and then
  recommend a set of Public Policies?
• To start the discussions should
• The Federal Government establish a timeframe and
  fund the following research areas
• Invest to maintain the current performance growth
  in HPC?
• Invest in new cooling technologies to improve
  efficiencies?
• Invest in lower power, higher performance
  processor(s)?
• Invest in new material research for chips?
• HPC data centers will have to become more
  accountable for power and cooling consumption

22
Phase 2 StudyStudy of Power Cooling of
Enterprise Data Centers
23
Study 2 Objectives

• "Potential Solutions to Address and Improve Power
  and Cooling for HPC Servers
• Study to explore data center power and cooling
  technologies
• Practices, and future plans/solutions
• Exploring energy efficiency management
  toolsbeing used today and emerging solutions.
• The study will focus on new products and
  solutions
• Near term horizon, that can reasonably expect to
  be able within the next 1 to 2 years.
• Study will include surveys of providers of
  solutions and HPC vendors, in addition it will
  include a "testing" of how likely each solution
  will actually become available in the market in
  the near term.

24
Study 2 Objective (contd)

• The study intends to leverage the recent
  IDC/Avetec study
• Identified the key strengths and shortcomings of
  current cooling approaches
• Identify those that are under research by
  suppliers
• As an important side benefit, the study will also
  document user pain points and concerns about
  future HPC data center cooling solutions.

25
Potential Impact and Outcomes

• Create a policy document that will assist US
  government and manufacturers to develop new
  products and perhaps laws to enhance overarching
  computer and communications facilities in
  conjunction with improvements in power and
  cooling demands.
• Compare and contrast solutions and approaches
  among the communities. Approaches that work
  across communities have strong potential to be
  leveraged for economies-of-scale that can lead to
  lower costs.
• Recommend a set of test tools (hardware and
  software) and a methodology to assist centers in
  self-assessment.

26
Future Solutions and Technologies

• Potential technology solutions
• Advanced power management in hardware
• Dynamic network power management
• New data storage technologies
• Hardened equipment
• Novel computing architectures
• Game-Changing Technologies
• All-optical networks
• Superconducting components
• Nano-electronic circuitry
• Eliminate voltage conversion steps
• Liquid cooling of hardware components
• Transition to DC operations

27
Conclusion and Recommendations

• Data Centers
• Proactively work with vendors, researchers and
  the HPC community to identify and pursue
  breakthrough power and cooling technologies.
• If not currently monitoring power and cooling
  consumption and cost, implement an approach and
  methodology to measure and track consumption.
• Identify and implement methods to reduce
  consumption.
• Install new air conditioning and power systems
  that reduce power consumption.
• Establish comprehensive performance metrics to
  track and report results of new initiatives.
• Investigate enhancing cooling methodologies
  through the use of liquid technology.

28
Conclusions and Recommendations

• Vendors
• Develop technologies that will drive down power
  and cooling demands by using technologies that
  dynamically power up processors on demand.
• Technology vendors need to design their systems
  requirements to address end-to-end data center
  infrastructure in order to develop the most
  efficient center possible.