Austin
Data
Center
Project
Feasibility
Study


TACC‐SECO
Final
Feasibility
Report

CM1001  
 
Prepared by 
Dan Stanzione
 TACC­SECO Feasibility Report 

Executive
Summary

 
The SECO‐TACC project was designed to explore the feasibility of a shared datacenter 
facility for Texas that could provide substantial cost and green benefits to the individual 
participants as well to the state as a whole.  
 
The Texas Advaned Computing Center (TACC) team has taken a comprehensive approach 
to exploration of a shared datacenter, looking in equal parts at technology for both green 
and shared datacenters, as well as the partnership and trust building required for a shared 
datacenter plan.  
 
The conclusions of this report are as follows:  
• It is feasible for a shared datacenter facility to be constructed in Texas that 
would involve the universities, state agencies, and corporate customers. 
• A large scale shared facility would likely provide significant cost and energy 
savings, as well as be an enabler for further economic development. 
• There are a number of barriers that would prevent full adoption of this 
facility in ways that would maximize power efficiency; nonetheless, the 
adoption that would happen would still bring substantial benefits.    
• Recent technological advances make a large scale shared facility possible at 
dramatically higher power efficiencies than most existing datacenters.  
• Barriers to shared‐system adoption are technical, legal, and psychological. 
• Cloud computing trends in the private sector will continue to lower the 
technical and psychological barriers over the next several years. 
• Any shared datacenter facility must pay special attention to compliance with 
HIPAA, FERPA, and other data protection standards. 
• Rather than offering fractional servers, offering secure hosted services may 
be a more successful approach.  
• Shared datacenter and shared service models are already gaining traction 
among in‐state universities, and TACC is facilitating some of this. 

Page 2 of 31
 Table
of
Contents

Executive Summary............................................................................................................................................ 2 
Table of Contents................................................................................................................................................. 3 
1.0 Introduction: Vision of a high efficiency, shared datacenter for Texas................................ 4 
2.0 High Efficiency Datacenter Design ....................................................................................................... 5 
2.1 Experimental partnership with Green Revolution Cooling.................................................. 6 
3.0 Datacenter Partnerships.........................................................................................................................10 
3.1 Exploration of partnerships with Switch Technologies.......................................................10 
3.2 Proposals to potential datacenter partners ..............................................................................11 
4.0 Cloud/Virtualization technology evaluation .................................................................................12 
5.0 Practical issues in a shared, high efficiency datacenter............................................................13 
6.0 Final Recommendations .........................................................................................................................16 
Appendix A – New datacenters around the world..............................................................................19 
Appendix B – Comparison of rack top to chilled water data centers (commissioned as 
part of a datacenter design for UT from HMG Associates)..............................................................25 
Appendix C: Cost estimates for Datacenter ...........................................................................................26 
C.1  Site Plan ...................................................................................................................................................26 
C.2  Preliminary Cost Estimates ..........................................................................................................26 
C.3  Preliminary Cost Estimates (for the base work included above) ................................28 
C.4 Projected Schedule...............................................................................................................................31 

Page 3 of 31
1.0
Introduction:
Vision
of
a
high
efficiency,
shared
datacenter
for
Texas



The SECO‐TACC project was designed to explore the feasibility of a shared datacenter 
facility for Texas that could provide substantial cost and green benefits to the individual 
participants as well to the state as a whole.  
 
The TACC team took a comprehensive approach to exploration of a shared datacenter, 
looking in equal parts at technology for both green and shared datacenters, as well as the 
partnership and trust building required for a shared datacenter plan.  
 
The technology evaluation activities included evaluation of various cooling technologies, 
including both devices for raising efficiency commercially available today, such as rack 
door and rack top cooling units, and more experimental technologies, such as the 
evaluation of a mineral‐oil immersion technology.  Additional evaluation work was focused 
on technologies for allowing hardware sharing, specifically virtualization technologies such 
as Eucalyptus.  
 
The second set of activities focused in building partnerships, both with potential datacenter 
customers as well as potential providers, to advance a plan for producing a large scale, high 
efficiency, multi‐customer datacenter in Texas.  
 
This report summarizes the activities in both categories that took place during the course 
of the project. Partnership building  activities included:  
• Exploration of partnerships with Switch. 
• Proposals to potential datacenter partners 
• Datacenter tours 
• Analysis of prospects for shared datacenters. 
Technology evaluation activities summarized include:  
• High Efficiency Datacenter Design 
• Experimental partnership with Green Revolution Cooling 
• Cloud/Virtualization technology evaluation 
• Survey of other significant datacenter projects in academia and industry 
(appendix).  
 






Page 4 of 31
2.0
High
Efficiency
Datacenter
Design



TACC has continued investigation into the design of more efficient datacenters through the 
use of the best available off the shelf cooling technologies.  The current TACC datacenter, 
constructed 3 years ago, already incorporates technologies beyond those used in most 
conventional datacenters to achieve higher efficiency at high density.  While employing 
racks using more than 30kw each, the TACC datacenter uses in‐row coolers (IRCs) from 
APC to bring chilled water closer to the racks, and an enclosed hot‐aisle technique to 
further enhance the effectiveness of the IRCs.  Comparisons to more traditional datacenter 
designs imply that employing the IRCs reduces total cooling power for the datacenter by 
around 15%.  The TACC datacenter is considered a model for modern datacenter efficiency, 
and during the course of the project, TACC staff have made a number of public 
presentations  describing our approach to high density datacenters, and seeking industrial 
partners in constructing a new one.  
 
TACC continues to investigate alternative technologies commercially available to 
determine if more efficient technologies could be used.  In a recent exercise with an design 
firm, the current commercial offerings in rack top and rack door chilling were compared 
with new generation IRC technology, as well as investigations of alternate schemes like 
importing outside air.  The relatively high temperature and humidity in central Texas ruled 
out the outside air options (though this approach may work well in west Texas, perhaps El 
Paso, where less dehumidification would be required).   The final design exercise came 
down to a decision between rack door chilling units and IRCs.  Ultimately, IRCs were 
deemed to still be the best option for raising cooling efficiency.  Appendix B contains the 
engineer’s final analysis on these two options.  
 
TACC continues to track the latest ASHRAE standards for operating conditions for new 
computer equipment.  Next generation computer hardware is expected to operate 
effectively at significantly higher temperatures, and across a higher range of humidity 
values.   This new operating band may make the use of outside air a more attractive option.  
The experience of TACC staff on current systems, however, shows that with current 
hardware even modest increases in operating temperature generates a significant increase 
in the rate of disk drive failure.  TACC is evaluating the potential of running cluster nodes 
without disk drives in each node to mitigate this weakness (the Ranger cluster compute 
nodes use a solid state storage device, or SSD, instead of a spinning disk).   The Ranger 
experience has shown that solid state drives are a viable option at scale for reducing disk 
drive temperature related failures, though the small size and limited write cycle current 
SSDs were unpopular with the systems administration staff.  

Page 5 of 31
2.1
Experimental
partnership
with
Green
Revolution
Cooling

 
In addition to evaluating off the shelf technology, TACC investigated experimental 
technologies to boost datacenter efficiency that had not yet reached the commercial 
marketplace (although the primary one was put on the market during the course of the 
project).    
 
The most promising of these investigations has led to the TACC partnership with Green 
Revolution Cooling (GR).  The GR team is designing a product that allows computing 
equipment to operate while immersed in mineral oil.  The high heat conductance of mineral 
oil makes it dramatically simpler to remove heat from the equipment, dramatically 
reducing the cooling infrastructure required.  The PUE (power use efficiency) of a typical 
current generation commercial datacenter is often around 1.4, meaning in addition to the 
power for the systems, an additional 40% power is required to run cooling systems to 
remove the heat (in many older datacenters, particularly small installations, the PUE can be 
as high as 2.0).  TACC finds the GR technology particularly intriguing as the potential exists 
to reduce the PUE below 1.0.   While some power is still required for cooling, this 
innovative approach allows the system fans within each computer to be shut off or 
removed.  This reduces the actual power to the computing equipment by 10‐15%.  The 
total power required for cooling is only 3‐5% with the GR approach.  This means GR cooled 
systems may require less power than the computers running with zero cooling via 
conventional means. 
 
TACC has entered into a partnership with GR, which resulted in the first deployed 
production prototype of this technology at TACC for evaluation in April of the project year.     
Between May and the end of the project, TACC operated this system with a variety of 
hardware configurations, and measured both power required to cool the system, and the 
reliability of both the cooling infrastructure, and the computing equipment immersed in 
the mineral oil.    A photo of the equipment installation at TACC is shown in figure 1.    
 
Not shown in the picture is a small evaporative cooling unit, which is attached via a pipe to 
the heat exchanger to the left of the rack.   The basic operation of the system is to immerse 
the servers vertically in the rack, and create a slow flow of oil through the servers from 
bottom to top.  The oil is them pumped through the heat exchanger, where excess heat is 
transferred to water.   The water is then cooled evaporatively outside (or through any 
other cooling source).    
 
The energy savings come from several sources.   The primary advantage derives from the 
fact that mineral oil is 1200 times more effective as a heat conductor than air.   As a result 
of this, energy can be removed from the systems much more efficiently.  In an air cooled 
solution, the ambient air in the datacenter is typically cooled to 65‐80 degrees F in order to 
keep the actual processor chip inside the computers running below 140 degrees F.   
Because the mineral oil is such a superior conductor of heat, in this solution the ambient 
temperature of the oil can be raised to 100‐105 degrees F.    In an air cooled data center, 

Page 6 of 31
water is normally chilled to 45 degrees in order to support the air conditioning units, 
requiring significant energy. In the GR solution, the water to supply the heat exchanger 
only needs to be cooled to around 90 degrees.   Simply running the water through a small 
evaporative unit and cooling it with outside air was sufficient to cool the oil throughout a 
Texas summer, running 24 hours a day.   A secondary source of power savings is that all 
fans can be removed from the servers.  The fans are used merely to increase the rate of 
airflow across critical components, to improve heat conduction.  The normal heat 
conduction of mineral oil is sufficient to not require acceleration. The removal of fans 
further reduces the base operating power of the servers by 5‐15% (depending on load), in 
addition to the savings in external cooling power.  
 
 

 
Figure 1: The Green Revolution Prototype Cooling Rack at TACC.  The system is able to efficiently cool computing 
equipment even when exposed to outdoor air. 

 
With the GR solution, typical commodity computer equipment runs fully submersed in the 
computer equipment with only a few simple modifications.   For the TACC evaluations, we 
used primarily Dell servers, with a few servers from other vendors.  Typical server 
modification for insertion in the oil took less than 10 minutes.   When compared to other 
liquid‐cooled solutions from other vendors, a substantial advantage of the GR approach is 
the ability to use commodity‐built hardware from major vendors.   For the bulk of the 
evaluation at TACC, we ran a load of approximately 8KW in the mineral oil continuously for 

Page 7 of 31
4 months.   During this period, the measured total power for cooling (pumps, heat 
exchanger, evaporative cooling unit) varied from 60 watts to 240 watts (depending on 
pump rate and whether or not the cooling tower was active).   The average power was 
below 150 watts, or less than 2% of the total load.  
 
The implications of this level of power savings are staggering.  Consider the case of the 
TACC data center that houses the Ranger supercomputer.  This is a modern datacenter, 
using in‐rack chilling technology and built for high efficiency less than 4 years ago.   With a 
PUE between 1.3 and 1.4, it is at least as good as any other modern large scale facility, and 
vastly superior to the typical “data closet” used in most facilities in terms of efficiency.   
Despite this efficiency, this datacenter uses approximately 1 million watts at all times for 
cooling, or about 8.7 million kw/h per year.   The costs of generating this cooling are about 
$400,000 annually, and require the equivalent energy output of about 3,000 tons of coal.    
A switch to the GR technology would reduce this consumption by 98%, and that is before 
taking into account additional savings from removing the server fans!   
 
When you consider that it is believed that 1‐3% of all US power consumption is already 
spent on datacenters, and that between 1/3‐1/2 of that power is used for cooling, the 
enormous potential for savings with this technology become clear. GR Cooling estimates 
the savings for a “typical” rack (20KW) to be $78,000 over the course of a 10 year 
datacenter life.    
 
Given the success of the trial and the tremendous potential of this technology, we moved 
our evaluation on to considering other practical considerations of deploying this 
technology in a production‐scale datacenter.      
 
The first concern was reliability.   This can be separated into two problems, the impact of 
reliability on the computing equipment itself from mineral oil immersion, and the 
reliability of the cooling infrastructure.    Given that in this pilot project we performed only 
a 4 month evaluation with a single cabinet and a small number of servers, the reliability 
results must be considered preliminary at best.  However, the results so far are extremely 
encouraging.   We suffered no server failures during our evaluation, and prolonged 
exposure to the oil seemed to have no ill effects.  We believe that server reliability will 
actually be slightly enhanced through this approach.   The removal of the fans should 
reduce failures.  The fans comprise most of the moving parts in a server (disk drives are the 
other), and therefore are a significant source of failures in normal operation.  There 
removal can only enhance reliability.   Further, there is significant evidence in large scale 
power systems that mineral oil immersion can improve component reliability (mineral oil 
cooling is used in transformers and other components in the power grid).  As a better 
electrical insulator than air, mineral oil immersion should reduce micro‐arcs, small sparks 
that normally corrode electrical connectors.  This should provide another slight boost in 
reliability.      
 
During the course of our evaluation, we had only a single failure in the cooling 
infrastructure.   The root cause of the failure was faulty circuit wiring, with a shared ground 

Page 8 of 31
between multiple breakers causing a current overload.   While this particular failure is 
unlikely to reoccur, this does point out the need for redundancy in the infrastructure.  
While our prototype has a backup pump and backup heat exchanger, it has only one cooling 
tower, and a shared power feed to the primary and redundant pumps.   A production 
solution would require separate power feeds to the backup systems, and redundancy in the 
tower.    This is not an onerous requirement.   Our current datacenter design has 7 
independent CRAC (Computer Room Air Conditioner) units and 3 independent chilling 
plants, to provide sufficient redundancy to survive failures.  The GR infrastructure would 
simply need similar levels of redundancy.   
 
Another concern in putting this technology into production is density, i.e., could a 
datacenter layout built from this technology at scale support the same number of servers as 
an air cooled datacenter in the same square footage count?   We have put together a 
comparative projection in conjunction with GR Cooling.   The fundamental issue is that the 
oil‐cooled rack is the equivalent capacity of a normal rack, but while a normal rack 
provides 42U (42 “rack units”) of server space stacked vertically, the GR rack provides 
those 42 units side‐by‐side.   So, for a single cabinet, the GR unit provides a larger physical 
footprint.   However, for a full size data center layout, there are a number of space 
advantages.    
 
A traditional air cooled rack requires an aisle to be left at both the front and the back of the 
rack, of suitable size both to support sufficient airflow and to allow for maintenance and 
removal of servers.  The typical size of this aisle is 4’.  With the GR solution, servers are 
removed from the top of the rack, so two cabinets can be placed back‐to‐back, removing the 
need for an aisle.   While the GR racks require additional floor space for the pump 
infrastructure, they remove the need for CRAC units (which are larger) to be placed on the 
floor.     The hypothetical layout for a 24 rack GR‐based datacenter requires 38 square feet 
per cabinet.   The current Ranger datacenter supports 100 racks in approximately 4,000 
square feet, or roughly the same amount.   So, an equivalent number of rack units can be 
supported in the same physical space.   The hypothetical layout of our mineral oil cooled 
datacenter is shown in figure 2. 
 
Perhaps a more important measure of density is the total wattage of computing capacity 
that can be supported in a given sized datacenter.   Current racks at TACC draw 30KW.   
The limit of air cooling schemes may fall at between 40‐60KW per rack.   Though we do not 
yet have computers at this density to test the hypothesis, we believe 100kw per rack is 
possible with the GR solution.  Such dense computing hardware should become available in 
the next 4 years.    So, while a GR‐style datacenter should support an equal number of racks 
as an air‐cooled datacenter, the quantity of equipment per rack should be substantially 
higher with the oil‐based solution. 
 

Page 9 of 31
  
Figure 2: Hypothetical layout of 24 cabinet mineral oil­cooled datacenter, with supporting infrastructure. 

 
A final concern is the environmental and safety impacts of using mineral oil in the 
datacenter at large scale.   The mineral oil used in this solution is non‐toxic.   In fact, grades 
of mineral oil can be used within the racks that are considered fit for human consumption.   
Handling and exposure provides no significant risks.  This grade of mineral oil also poses 
little fire hazard.  While it will ignite in extreme heat, under normal conditions a lit match 
can be dropped into the oil and will simply be extinguished.  
 
TACC finds this cooling option to be particularly promising and exciting.  The combination 
of extreme energy savings, strong reliability, and the ability to simply adapt commodity 
servers makes this a very attractive technology for a future green datacenter.    TACC 
continues to evaluate this technology and explore additional funding in conjunction with 
GR to support this concept.  
 

3.0
Datacenter
Partnerships

3.1
Exploration
of
partnerships
with
Switch
Technologies

 
Switch Technologies operates the SuperNAP datacenter in Las Vegas, a 150MW, 400,000 
square foot datacenter which is among the largest and most efficient hosting centers in the 
world.  Switch is seeking to build additional facilities of approximately 1 million square feet 
at several additional strategic sites in the United States.   TACC has forged a coalition with 
UT‐Austin, UT‐System, and Austin community leaders to work with Switch to locate a new 
large scale datacenter in Central Texas.  Representatives of Switch have made several trips 
to the area to meet with the Chamber of Commerce, local power companies, and to visit 
potential sites.   TACC staff have also arranged visits for university officials to Switch’s Las 
Vegas facility.   
 

Page 10 of 31
As the ultimate goal of this project is to determine the feasibility of constructing a large 
scale, shared datacenter in Texas, the relationship with Switch could be a critical 
component.  The ultimate goal of this investigation is to forge a partnership with Switch 
whereby they would invest in the construction and operation of this facility, and the 
university and other state entities would enter into a long term Memorandum of 
Understanding (MOU).  The addition of corporate customers Switch could attract (some of 
which TACC could facilitate) would defray the cost to the state for the use of this facility.  
Negotiations for an MOU with Switch and a possible site in Austin are ongoing.  
 
The TACC team also continues to investigate other potential large scale datacenter options 
as well.  As part of this project, we also visited the Citibank datacenter in Georgetown, as 
well as doing site explorations with the Domain, the Met Center, and Montopolis to explore 
power and cooling options.  
 

3.2
Proposals
to
potential
datacenter
partners

 
A non‐technical issue which will be key to the ultimate success of any Texas datacenter 
proposal will be the engagement of additional occupants of the facility beyond TACC.   The 
TACC team has engaged in a number of activities to gauge the interest level of other 
potential customers in such a shared facility.   Discussions thus far include: 
• Presentations to TACC’s Science and Technology Affiliates for Research  
(STAR) industrial partners, including Chevron BP, and Shell on the concept of 
a high efficiency shared facility. 
• Discussion with other Texas state institutions about the use of both a shared 
facility and shared systems through the High Performance Computing Across 
Texas (HiPCAT) consortium.    
• Meetings with UT‐System about partnerships with the medical institutions to 
share in a facility. 
 
As a result of these discussions, several concrete partnerships have emerged that can be 
leveraged in a future datacenter investment.   One partnership was formed between TACC 
and both Texas A&M and Texas Tech universities.   Both Tech and A&M agreed to invest 
$500k in the next TACC supercomputer, recognizing the increased capabilities and 
economies of scale of investing in a large shared system.  While this partnership focuses on 
systems rather than datacenter space, it represents both an unprecedented level of 
collaboration, and a foundation for future collaborations on shared facilities as well as 
shared systems. 
 
A second partnership of note is a relationship to back up UT System data in the TACC 
datacenter.  Part of this relationship will be the recognition of UT‐Austin and TACC as a UT‐
System “shared datacenter”; while not yet a statewide datacenter, and not yet a focus on 
green, this partnership will allow all 15 UT institutions to purchase datacenter services 
through TACC, a foundation for a larger statewide shared datacenter agreement.       
 

Page 11 of 31
In a related development, TACC is developing a proposal with the 15 institutions to develop 
shared, centralized data storage and computation systems.  This initiative would stop short 
of a shared datacenter (the partner institutions would receive allocations on a time shared 
computing cluster, and share space on a large disk system). While not yet funded, this 
proposal is another step in the evolution to a vision of a large scale, shared facility, rather 
than replicated facilities across the state.   
 
In meetings with industrial partners to gauge interest in buying into a shared datacenter 
facility (primarily the energy companies), we have found interest is high, though some 
concerns remain.  For instance, one large oil company finds our pricing on a shared facility 
involving comparable to their internal costs, and a recent large internal investment would 
preclude participation for the next 4‐5 years.  Nevertheless, we feel there would be a strong 
possibility of attracting at least a few corporate customers to be tenants should a statewide 
datacenter be constructed.  

4.0
Cloud/Virtualization
technology
evaluation

 
Technologies for cooling are only one part of the equation for building a high efficiency 
datacenter.   The other technologies to boost datacenter efficiencies are those which allow 
multiple users to actually share hardware, reducing the total number of computers that 
must be deployed.   While there is currently much resistance to this idea, there is significant 
precedent; in the early days of mainframe computing, all computers were timeshared, and 
systems were so expensive that companies routinely leased time on shared systems.   The 
TACC supercomputers have continued to share this philosophy, with the current systems 
supporting more than 400 projects and maintaining more than 90% utilization; much 
higher utilization than 400 small clusters would maintain, using dramatically less total 
hardware and power.  
 
While timesharing has been around since the beginning of computing, the recent rise of 
virtualization technologies have enabled a new class of system sharing.  While 
virtualization itself is an old idea, pioneered by IBM nearly 40 years ago, only recently has 
virtualization become practical on commodity, low‐cost servers.   Simply put, virtualization 
provides a user with the illusion of having their own server; however, this server is 
“virtual”, it exists inside of a real server, but may share it with other virtual machines, or it 
may move from one physical server to another.   In the last few years, virtual machine (VM) 
performance has improved to the point where for many applications, running in a VM is 
nearly indistinguishable from running on a physical system.  
 
To date, the project team has constructed a cluster of systems using the Xen VM software 
that are now used for a variety of information technology tasks at TACC.  Performance 
characterization of these systems has shown thus far that processor intensive tasks suffer 
little performance penalty, though Input/Output intensive applications may still see 20‐
30% degradation from VMs.   In general, VMs have proven robust enough for normal use.  
Indeed, virtualization is now the basis for a large number of commercial cloud and hosting 
options, proving the viability of this technology at scale.  

Page 12 of 31
 
Most recently, the project team has set up a new virtual cluster using the Eucalyptus open 
source software package.  Eucalyptus allows the creation of a “private cloud”, e.g., it 
manages a set of physical resources and dynamically schedules sets of virtual machines 
from a queue of requests.  The Eucalyptus software can run VM images compatible with 
Amazon’s Elastic Compute Cloud (ECC).   The current evaluation of Eucalytpus is to analyze 
the security model, i.e. to evaluate if the system will provide sufficient data protection to 
allow multiple customers to share virtual machines in the same physical systems without 
risk of exposing sensitive data.  While VMs are now robust enough for the workloads of 
many clients in a shared datacenter, the primary concern is protection of one client’s data 
in VMs on shared systems from other users.  This concern is discussed more fully in the 
next section.    
 
Our evaluation included Eucalyptus and the XenServer software.   A third viable, and 
perhaps the strongest offering in this space, is VMWare server.    Budgetary considerations 
precluded a thorough evaluation of VMWare.    The evaluations of Eucalyptus and 
XenServer included hands‐on experience with small clusters running multiple VMs, as well 
as interviews with the operators of the largest current installation of Eucalyptus, the 
Nimbus system at Argonne National Laboratories.   
 
The evaluation determined that Eucalyptus is useful for research and academic projects, 
but is not yet mature enough for large production installations.   While smaller sets of 
systems work well, at large scale, there are still many bugs at “edge conditions” as the set of 
VMS exhaust the resources of the physical servers.   XenServer seemed more bullet proof 
and stable, and offered more consistent and repeatable performance.   Anecdotal evidence 
implies VMWare ESX is perhaps the most robust, but at a very high cost.    
 
The VM software evaluation proved to us that virtualization at this point is a robust enough 
solution to provide server sharing for a set of systems where there is trust among users (i.e. 
within many groups in a large company).   However, security concerns with all of these 
products make us hesitant to recommend this solution to enable the sharing of servers 
between distinct customers who have a competitive relationship or lack of explicit trust.   
See the further discussion of this issue below in section 5.  




5.0
Practical
issues
in
a
shared,
high
efficiency
datacenter



 
Technically, there are very few barriers remaining to building a large scale, multi‐customer, 
shared server datacenter.  In fact, a number exist.  However, there are still a number of 
practical barriers to constructing a large datacenter that effectively implements sharing 
within a rack.   Most of these issues stem as much from perception as technical reality.  
 
The most fundamental issue is the issue of data security.  The perception exists that any 
data stored with a remote provider is inherently less secure than data stored at the owner’s 

Page 13 of 31
site.  This assertion is almost certainly untrue.  Maintaining data security is an ever more 
complex task.  The list of compromises of individual datasets by companies or government 
agencies is simply too long to even attempt to elaborate.   In this climate of growing 
complexity, there is a distinct advantage to having a large investment in security 
professionals.   While organizations still perceive that keeping data “in‐house” is the most 
secure method for storing data, the simple fact is that large, IT focused organizations are 
much more likely to be able to sustain the investment required to truly make data systems 
secure.  The rise of “cloud computing” providers are slowly changing this culture.  
Companies like Amazon.com and Google are capable of investing tens of millions quarterly 
into security for shared systems,  a larger investment than most universities, government 
agencies, or small to mid size companies can make in IT total.  The success of the cloud 
model is likely to slowly start changing this mindset towards “remote” becoming 
synonymous with “secure”.  
 
From a technical perspective, virtually all organizations have mastered the art of 
distributing secure access to their systems.  The widespread use of VPN (Virtual Private 
Network) technology has made it routine for companies to offer secure access across 
distributed worksites, or to employees outside the office or at home.  The move to a remote 
datacenter is simply making the psychological change that the one of the sites on the VPN is 
no longer in a company owned space.   The technical challenges for a remote datacenter or 
largely solved, at least for the case of dedicated servers living in a remote space.  For 
shared, multi‐client servers, more technical barriers remain, and the use of secure 
virtualization software is being investigated as part of this study.   While the technology 
exists to isolate users from each other effectively through shared systems, the weakness in 
current virtualization software is that the administrators of the shared systems can still 
access all of the machine images on the shared system.  So, either a framework of trust 
must be developed with providers, or current offerings must be limited to a “shared space” 
but not a “shared system” approach.    
 
This is perhaps not as large an issue as it would at first appear.  First of all, trust can likely 
be established between a large enough set of groups to allow a relatively large set of 
servers to be shared (for instance, all state agencies sharing servers operated through a 
state‐run virtual pool).   Second, most organizations require many servers.  The benefits of 
sharing within the group provide almost all of the available advantage through 
virtualization; sharing with another organization provides at best a marginal benefit.    
Consider for instance this example.  Suppose three departments choose to use 
virtualization to consolidate their servers.  Also suppose that each physical server can 
accommodate up to 8 virtual machines (in practice this is slightly more dynamic, but the 
principle holds).   Suppose department A has 30 servers, department B 35, and department 
C 20.    Let’s consider a “server” to be a physical machine consuming 500 Watts at nominal 
load. Figure 3 below lists the number of physical servers required by each department, and 
total, in each of three scenarios:  Physical servers, virtualization with no shared servers 
between departments, and full virtualization with shared systems.  
 
 

Page 14 of 31
 
Scenario  Physical  Virtualization  Virtualization, 
Servers  servers shared 
between 
departments 
Dept. A  30  4  3.75 
Dept. B  35  5  4.325 
Dept. C  20  3  2.5 
Total  85 (42.5KW)  12  (6KW)  11 (5.5KW) 
Figure 3: Comparison of server count and power consumption for 3 departments using physical servers, 8 way 
virtualization, and virtualization with sharing between departments. 

In this scenario, simply employing virtualization within a department reduces total power 
from 42.5KW to 6KW, a reduction of total power by 86%.    The additional step of sharing 
servers between departments in this case saves only an additional 0.5KW, less than 2% of 
the original power (though a more significant 8% of the virtualized power).   A datacenter 
could be made that was substantially “green” simply by providing virtual pools between 
groups that can establish a sufficient level of administrative trust.  
 
While the technical challenges of a shared datacenter are largely solved, a more complex 
set of problems revolve around the legal and regulatory framework regarding the retention 
and privacy of data.  If an organization stores, for instance, data about medical patients, the 
datacenter and systems housing it must be in compliance with the federal HIPAA laws.   
Data such as customer names and credit card numbers has a less rigorous legal standard, 
but still carries significant civil liability risk if disclosed.   The legal and regulatory 
frameworks surrounding privacy issues add substantial complications for the operation 
both of shared datacenters and shared systems.   Many organizations currently interpret 
HIPAA, for instance, to essentially require all HIPAA systems to reside in an isolated, 
physically secure datacenter.    
 
However, this typically seems to involve an overly conservative interpretation of the 
regulation.  With the increased emphasis on electronic medical records, cloud services 
providers such as DiCom Grid (http://www.dicomcourier.com/about/) have begun to 
demonstrate HIPAA compliant services delivered over the internet from a shared, secure 
datacenter.  DiCom’s ImageCare platform allows small health care facilities to store images 
remotely at DiCom’s datacenter while remaining compliant. A much more rigorous 
software approach to encryption and data transfer is required, however, shared facilities 
are still possible.   For large hosting providers, typical solutions involve building secure 
“zones” within a shared datacenter, where separate network facilities can be provisioned, 
and separate access logs detailing who gains physical access to that zone can be maintained 
and audited.  
 
To summarize, security and comfort with the idea of “outsourced” systems remain 
significant barriers to a multi‐use datacenter, but the rise of cloud computing is changing 
the perception of these issues.  For a datacenter with shared physical space, there is 
primarily a social problem, secondarily a legal/compliance problem that is probably 

Page 15 of 31
manageable, and very few technology problems remains.  For a facility with shared 
systems, there are more significant but likely solvable technology problems, but perhaps 
insurmountable legal and regulatory problems.   Fortunately, large cloud providers with 
significant political influence, including Google, Microsoft, and Amazon may be able to 
affect the relevant regulations and laws over time.  

6.0
Final
Recommendations


 
The primary conclusions of this project is it is clear that (1) there are significant economies 
of scale in large datacenters, (2) that datacenters can be built that are dramatically more 
efficient than typical datacenters that exist today, and (3) that it is practical at the very least 
to share the physical facilities between many customers.  At this point, it is less clear that 
sharing at the individual server level is practical – while technically feasible, it is unclear if 
most customers at this point would be willing to accept the perceived security risk.   We 
believe firmly that the psychological, regulatory, and legal barriers that have in the past 
inhibited large scale shared facilities are already surmountable, and are becoming 
progressively lower as market forces reduce them.     It is also clear that the gap in 
efficiency between a large scale, professionally run datacenter, and a small, in‐house 
datacenter has grown substantially, and continues to grow.   Consider that a cutting edge 
green datacenter may now involve a many megawatt facility with redundant high voltage 
lines, a mineral oil cooled infrastructure on an industrial floor, a suite of virtualization 
services to be monitored and maintained, a set of information security experts, a set of 
compliance experts capable of generating secure network segments and tunnels for routing 
sensitive data subject to regulatory concerns, and 24 hours operations.    How many 
individual organizations have it within the scope of their budget to achieve this level of 
efficiency, and provide the many kinds of expertise required for reliable, secure 
operations?  How many organizations within the state should replicate this expensive 
functionality?  
 
This study has left little doubt that if constructed, a statewide datacenter could be made 
substantially more efficient than the many small existing datacenters in the state.   It would 
be adopted across the many universities of the state, by industrial partners, and hopefully 
by state agencies.   Collaborations required to do this are forming in an ad hoc way even 
without a central facility, as shown by TACC’s new partnerships with UT System, Texas 
A&M, and Texas Tech.     
 
The fundamental barrier at this point is initial capital; which state entity will bear the 
budgetary load of initial construction of such a facility?    TACC is eager to work with other 
entities throughout the state and the legislature to see such a project brought to fruition. 
 
As part of this project, TACC commissioned a cost estimate and design for construction of a 
complete datacenter at UT‐Austin’s Pickle Research Campus. The summary of costs is 
presented in Appendix C.  A finding of this study was that the campus would require 
substantial infrastructure upgrades to support a datacenter at the scale proposed, perhaps 
driving up construction costs by as much as $24M.  An alternate location with more 

Page 16 of 31
available electrical power infrastructure would substantially reduce construction costs. 
TACC is pursuing possible alternate sites.  
 

Page 17 of 31
 
 

Page 18 of 31
Appendix
A
–
New
datacenters
around
the
world


1. University Data Centers

Data Centers opened within the last 12 months or currently being built with a focus on “green”
technologies.

1.1 Indiana University Data Center

Cost: $32.7 million
Data Center space: 11,000 sq ft
Total Facility Size: 82,700 sq ft
Green Features: Yes
Website: http://it.iu.edu/datacenter/index.php

Highlights:
• $32.7 million funded by Academic Facilities Bonds ($18.3 million), Infrastructure
Reserves ($8.4 million) and Capital Projects/Land Acquisition Reserves ($6 million).
• Single story bunker style design
• A ten- to twelve-foot-high berm around most of the building. This berm, planted with
native grasses and drought-resistant plants, improves insulation and reduces heat gain on
exterior walls and eliminates the need for a potable water irrigation system.
• 2 - 1.5MW generators with a rough in for a third expansion unit
• 2200 (1100 now and 1100 future) Tons of cooling via 4 chillers (3 now and 1future) and
4 cooling towers (2 now and 2 future) rated to withstand 140MPHwinds
• 2 total flywheels rated at 750kva/675kw UPS with flywheel energy storage providing
appr. 20 sec of ride through at full load with provisions for future third unit. UPS systems
are paralleled in an N+1 arrangement.
• 1 UPS rated at 500kva with battery energy storage providing 8 minutes of ride through at
full load
• 2x1800A @ -48V DC power system for MDF room
• 1500 sqft of raised floor for MDF room
• 3 Primary machine room pods
Complete list specifications: http://it.iu.edu/datacenter/docs/IU_Data_Center_Facts.pdf

1.2 Syracuse University Green Data Center (IBM, NY State Partners)

Cost: $12.4 million

Data Center Space: 6,000 sq ft
Total Facility Size: 12,000 sq ft
Green Features: Yes
Website: http://www.syr.edu/greendatacenter/

Page 19 of 31
Highlights:
• IBM provided more than $5 million in equipment, design services and support to the
GDC project, including supplying the power generation equipment, IBM BladeCenter,
IBM Power 575 and IBM z10 servers, and a DS8300 storage device.
• The New York State Energy Research and Development Authority (NYSERDA)
contributed $2 million to the project.
• Constructed in accordance with LEED “Green Building” Principles
• The SU GDC features an on-site electrical tri-generation system that uses natural gas-
fueled microturbines to generate all the electricity for the center and cooling for the
computer servers. The center will be able to operate completely off-grid.
• IBM and SU created a liquid cooling system that uses double-effect absorption chillers to
convert the exhaust heat from the microturbines into chilled water to cool the data
center’s servers and the cooling needs of an adjacent building. Server racks incorporate
“cooling doors” that use chilled water to remove heat from each rack more efficiently
than conventional room-cooling methods. Sensors will monitor server temperatures and
usage to tailor the amount of cooling delivered to each server—further improving
efficiency.
• The GDC project also incorporates a direct current (DC) power distribution system.

SU Green Data Center facts: http://syr.edu/greendatacenter/GDC_facts.pdf

1.3 University of Illinois / NCSA - National Petascale Computing Facility

Cost: unknown
Data Center Space: 20,000 sq ft
Total Facility Size: 88,000 sq ft
Green Features: Yes
Website: http://www.ncsa.illinois.edu/AboutUs/Facilities/pcf.html

Highlights:
• The facility is scheduled to be operational June 1, 2010
• Data center will house the Blue Waters sustained-petaflop supercomputer and other
computing, networking, and data systems
• NPCF will achieve at least LEED Gold certification, a benchmark for the design,
construction, and operation of green buildings.  
• NPCF's forecasted power usage effectiveness (PUE) rating is an impressive 1.1 to 1.2,
while a typical data center rating is 1.4. PUE is determined by dividing the amount of
power entering a data center by the power used to run the computer infrastructure within
it, so efficiency is greater as the quotient decreases toward 1.
• The Blue Waters system is completely water-cooled, which (according to data from IBM)
reduces energy requirements about 40 percent compared to air cooling.
• Three on-site cooling towers will provide water chilled by Mother Nature about 70
percent of the year.

Page 20 of 31
 • Power conversion losses will be reduced by running 480 volt AC power to compute
systems.
• The facility will operate continually at the high end of the American Society of Heating,
Refrigerating and Air-Conditioning Engineers standards, meaning the data center will not
be overcooled. Equipment must be able to operate with a 65F inlet water temperature and
a 78F inlet air temperature.
Interview with IBM Fellow Ed Seminaro, chief architect for Power HPC servers at IBM, about
this synergy as well as some of the unique aspects of the Blue Waters project:
http://www.ncsa.illinois.edu/News/Stories/Seminaro/

1.4 Princeton University HPC Research Center

Cost: unknown
Data Center Space: sq ft
Total Facility Size: 40,000 sq ft
Green Features: unknown
Highlights:
On January 20, 2010 Princeton University announced plans to build a facility to house its high-
performance computing research systems on the Forrestal Campus in Plainsboro about three
miles north of the main campus. The High-Performance Computing Research Center would be
located on the University's property and would serve as home of TIGRESS -- the Terascale
Infrastructure for Groundbreaking Research in Engineering and Science Center.

The new facility would have approximately 40,000 square feet and would comprise three
functional components: a computing area; an electrical and mechanical support area; and a small
office/support area. The two-story building would be about 50 feet high.

If approvals and construction proceed as planned, the facility would be operational in 2011 and
would be staffed by three people. It is expected to support the University's program needs
through at least 2017. The facility is sited to allow for future expansion; a second phase of
construction potentially could double the square footage.
http://www.princeton.edu/main/news/archive/S26/39/58I51/index.xml?section=topstories

Page 21 of 31
2. Corporate Data Centers
Recently opened or announced “green” corporate data centers.

2.1 Green House Data

Cheyenne, WY

Highlights:
• GHD is a 10,000 sq. ft. data center that is powered entirely through renewable wind
energy.
• GHD operates its facility at approximately 40-60% greater energy efficiency than the
average data center.
• The data center leverages the following attributes to gain the efficiencies:
o Air-Side Economizers – Free cooling from Cheyenne's average annual
temperatures of 45.6 degrees.
o Hot-Isle Heat Containment – Maximizing cooling efficiency by enclosing the hot-
isle and capturing or exhausting heat as our state of the art control system
determines.
o Modular Scalable Data Center – matching maximum efficiencies without over
building and waste.
o Efficient Floor Layout and Design – aligning hot aisle/cold aisles and redefining
the cage space concept.
o Highly Efficient IT Equipment – spin down disk technology and servers with the
highest power to performance ratios.
o Virtualization – cloud computing environment that reduces energy waste from
idle servers and IT equipment.
• Power – GHD has built an N+1 electrical power infrastructure that delivers power to its
customers in a true ‘A’ and ‘B’ power configuration. The facility has the capability of
providing up to 12.5kW of power to each data cabinet. The facility receives its power
from the local power company via a redundantly switched substation. Our internal power
infrastructure includes ATS, Generator and UPS protection to each rack.

Facility Overview: http://www.greenhousedata.com/greening-the-data-center/facility/

2.2 HP Wynyard Data Center

England
Highlights:
• Operational February 2010; 360,000 sq ft
• First ever wind- cooled data center
• Cooling: A seven-foot wide low-velocity fan is the entry point for the key ingredient on
HP’s innovative new data center in Wynyard, England: cool fresh air from the North Sea.
• 15 Foot Plenum: The large fans bring fresh air into perhaps the most innovative feature
of the HP Wynyard data center: a lower chamber that functions as a 15-foot high raised

Page 22 of 31
 floor. Inside this plenum, air is prepared for introduction into the IT equipment area.
When the outside air entering the facility is colder than needed, it is mixed with the warm
air generated by the IT equipment, which can be re-circulated from the upper floor into
the lower chamber.
• Filtering and Airflow: HP uses bag filters to filter the outside air before it enters the
equipment area. Once the air is filtered, it moves into the sub-floor plenum (which is
pressurized, just like a smaller raised floor) and flows upward through slotted vents
directly into the cold aisles of the data center, which are fully-enclosed by a cold-aisle
containment system. Capping the cold aisles in a “cool cubes” design allows the system
to operate with lower airflow rate than typical raised floors in an open hot aisle/cold aisle
configuration.
• Racks and Containment: HP uses white cabinets to house the servers at its Wynyard data
center, a design choice that can save energy, since the white surfaces reflect more light.
This helps illuminate the serve room, allowing HP to use less intense lighting. Another
energy-saving measure is the temperature in the contained cold aisle, which is maintained
at 24 degrees C (75.2F).
• Cabling and Power: The unique design of the HP Wynyard facility, with its large first-
floor plenum for cool air, also guides decisions regarding the placement of network
cabling, which is housed above the IT equipment. The UPS area is located in the rear of
the lower level of the data center, following the recent trend of segregating power and
mechanical equipment in galleries apart from the IT equipment. Heat from the UPS units
is evacuated through an overhead plenum and then vented to the outside of the building
along with waste heat from the server cabinets.
• At an average of 9 pence (11.7¢) per kWH, this design will save Wynyard approximately
£1m ($1.4m) per hall, which will deliver HP and its clients energy efficient computing
space with a carbon footprint of less than half of many of its competitors in the market.

Additional information and pictures: http://www.datacenterknowledge.com/inside-hps-green-

north-sea-data-center/

http://www.communities.hp.com/online/blogs/nextbigthingeds/archive/2010/02/12/first-wind-
cooled-data-center.aspx?jumpid=reg_R1002_USEN 

2.3 Facebook
Prineville, OR
Highlights:
• Announced January 21, 2010 - First company built data center; construction phase to last
12months.
• The 147,000 square foot data center will be designed to LEED Gold standards and is
expected to have a Power Usage Effectiveness (PUE) rating of 1.15.
• The data center will use evaporative cooling instead of a chiller system, continuing a
trend towards chiller-less data center and water conservation.

Page 23 of 31
 • The facility will also re-use excess heat expelled by servers, which will help heat office
space in the building, a strategy also being implemented by Telehouse and IBM.
• UPS design - The new design foregoes traditional uninterruptible power supply (UPS)
and power distribution units (PDUs) and adds a 12 volt battery to each server power
supply. This approach was pioneered by Google.
• Facebook Green Data Center Powered By Coal? – Recent reports suggest the announced
data center was not as green as previously thought. The company it will get its electricity
from is primarily powered by coal and not hydro power.
o Facebook responds to the criticism:
http://www.datacenterknowledge.com/archives/2010/02/17/facebook-responds-
on-coal-power-in-data-center/

Noteworthy Articles and Websites:

Google: “The World’s Most Efficient Data Centers”

http://www.datacenterknowledge.com/archives/2008/10/01/google-the-worlds-most-efficient-
data-centers/

Google Unveils Its Container Data Center

http://www.datacenterknowledge.com/archives/2009/04/01/google-unveils-its-container-data-
center/

Green Data Center News

http://www.greendatacenternews.org/

Data Center Journal – Facilities News

http://datacenterjournal.com/content/blogsection/6/41/

Data Center Knowledge

http://www.datacenterknowledge.com/

Page 24 of 31
Appendix
B
–
Comparison
of
rack
top
to
chilled
water
data
centers
(commissioned
as

part
of
a
datacenter
design
for
UT
from
HMG
Associates).



DATA CENTER COOLING OPTIONS

A. Chilled Water Cooling Doors

Sun Door 5200 System
• Adds 6.3 inches to blade chasis depth.
• Up to 6 times more efficient than traditional CRAC units.
• Up to 2 times more efficient than in-row cooling units (IRC).
• 30 kW per rack cooling capacity.
• Will still require supplemental cooling of 10 kW per rack or a total of 1 MW
(1,000 kW of additional cooling or 284 tons.) This would need to be in the form of
either IRCs with contained hot-isle containment of 10 additional CRAC units. If IRCs
were utilized, we must match the serve airflow and thus we would eliminate 20% of the
IRCs; in addition, if traditional CRACs were used, we would have mixing of all the room
air and the space temperature would need to be lowered and the efficiency of cooling is
greatly reduced. If the chilled door capacity met or exceeded that of the rack, then IRCs
would not be required and the costs would be lowered. However, the only
manufacturers we are familiar with have 15 kW, up to 30 kW doors, but not the 40 kW
required.

• Requires heat exchanger loop and higher chilled water supply (50°F-55°F
Supply).
• Does not control high temperatures in the cold-isle – requires supplemental
cooling.
• If chilled doors were utilized, the additional cost to the project would be about
$1,800,000 for the equipment only.

B. In Rack Coolers (IRC) placed in-row with the servers in a cold-isle contained hot aisle
configuration. This option is more efficient than the traditional CRACs and similar to
Ranger. We have asked APC to look at adding some additional IRCs in order to lower
the noise level, and evaluate if it is a cost effective option. If we used the IRCs we
would use traditional CRAC units for cooling the envelope and miscellaneous equipment
heat not in the in-row rack line up. The CRACs would also be able to cool
approximately 350 kW of additional miscellaneous equipment loads as currently
configured.

Recommendation: Based upon the fact the chilled water door capacity is 25% less than
actually needed, the cost benefits are dramatically reduced. Based upon this my
recommendation is to utilize the IRC option noted above and as discussed in our OPR
meetings of January 20 and January 21, 2010.

Page 25 of 31
Appendix
C:
Cost
estimates
for
Datacenter


The following is a detailed cost plan for a 5MW datacenter, constructed at the TACC facility,
prepared by an independent design firm. This design assumes a conventional raised floor
facility, with in-row cooler technology.

C.1

Site
Plan


 
C.2
 Preliminary
Cost
Estimates

 
Base Work, cost for the machine room expansion and the necessary satellite chilled 
water central plant (sized only to support the expansion project)       
              Base Costs     $32,000,000 
Additional Costs, proposed by Facilities Services (their Option 2):  
• Install 2 – 2,500 ton water cooled modular chillers at the site (which increases the 
capacity from 3,300 tons proposed above to 5,000 tons – Adds $2,000,000  
• Electrical upgrade from sub‐station to CP1 then TACC ‐ Adds $10,000,000 
• Austin Energy Transformer upgrade from 13 MW to 30 MW – Adds $12,000,000 
• Existing air cooled chillers would be shutdown and left in place for backup 

Page 26 of 31
 Additional  Costs    $24,000,000 

TACC Machine Room Expansion, with the Additional Costs 
  $56,000,000* 

*Note: the estimate is based on a construction start in 2010. A later start will need to 
add escalation. The Electrical upgrade costs have turned out to be required, so they are 
not optional. The total does not include the computer equipment and networking inside 
the machine room expansion, for which TACC is going to seek grant funding. 

Page 27 of 31
C.3
 Preliminary
Cost
Estimates
(for
the
base
work
included
above)


Page 28 of 31
  
 

Page 29 of 31
  
 

Page 30 of 31
  

C.4
Projected
Schedule

 
Austin energy projects a 24 month lead time for transformer upgrades at the Pickle 
Research Campus, so any construction at this site could not be completed before 
2013 if funding were immediately available.    
 
The transformer upgrade will be the dominant schedule factor in construction, so 
any project at Pickle would require at least 2 years from approval of funding.    
 
The outcome of this site plan and cost analysis indicates that the Pickle campus may 
be a poor choice for a new data center site.   Costs could be substantially reduced by 
moving to a site with available electrical infrastructure (such as Montopolis), or by 
moving out of UT facilities entirely.     
 
With the electrical upgrades removed, a 4 month time is estimated for completed 
design, followed by a 9 month period for facility construction.  
 

Page 31 of 31