Data centers have proliferated among U.S. federal agencies,
growing from 400 in 1998 to over 2000 by 2010. Alarmed by growing
costs and potential inefficiencies, the Obama Administration issued
a Presidential Memorandum directing all federal agencies to "adopt
a policy against expanding data centers beyond current levels, and
develop plans to consolidate and significantly reduce data centers
within 5 years.[1]" As a
result, Whitestone was tasked by a large agency to approximate the
facility savings that might accrue from the closure of a typical
data center.
We found that the range of possible cost savings
attributable to center closures is quite broad. Our estimates span
from $175 to $540 per square foot (SF) of computer room floor area,
varying by the amount of redundant equipment and computing
intensity. The precision of the estimates can be improved with
knowledge of a center's design features, as well as its size,
location, and levels of service.
Our definition of facility costs for data centers includes those
directly attributable to the center, such as energy and maintenance
and repair (M&R), and other operating costs-custodial, grounds,
management, pest control, etc.-that reflect standard tenant
services only indirectly related to the center. Not included are
the costs of the programmatic staff-operators, engineers, and other
technical staff-that run the computing activity. Unlike most
commercial data centers, many federal centers are small and occupy
only part of a multiuse building. Accordingly, our estimates assume
a size of 500 SF of computing area.
Cost were estimated for five types of data centers, as shown in
Table 1; four are categorized as Tier I through Tier IV centers,
according to the redundancy of mechanical equipment and the
independence of electrical distribution
paths.[2] A fifth "non tier" model is the
simplest version of a data center, with no raised flooring,
redundant systems or special power equipment. For each type of
center, we defined representative models based on component
inventories derived from case studies, and then verified by Jacobs
Engineering. These are summarized in the Attachment. Energy demand
was computed using an Energy Star
calculator.[3] Costs were generated using the
MARS Facility Cost Forecast System.[4]
Our estimates of energy costs do not vary for different tiers.
This is because adding redundant mechanical and electrical capacity
changes only the reliability of computing services, but not the
demand for computing. However, energy costs do vary by computing
intensity, as shown in Table 2.[5] Note that
the energy demand from computing (e.g. power required by servers,
drives, and switches) is only part of total center demand. Other
demand sources are from cooling and auxiliary equipment (e.g.
lights, UPS). Our highest annual cost estimate, $540 per computer
room SF, assumes a high level of computing intensity (substitute
$209.86 for the Tier IV model energy costs in Table
1).[6]
The M&R costs shown increasing by tier in Table 1 are a
function of the cyclical maintenance and replacement tasks
associated with the added mechanical and electrical equipment. The
reported costs are based on a 50-year average (a period long enough
to include the replacement of major cooling equipment).
Other operations costs are based on general tenant costs
calibrated for levels of use appropriate for data centers. For
example, security costs are based on systems and labor required for
the relatively high level of security common in federal buildings:
This includes "access control, system monitoring and intrusion
detection systems; stationed security guard and daily patrol." Note
that these costs increase by tier with the increase in "mechanical
room" space required by additional
equipment.[7]
Our estimates provide an approximate range of savings for an
individual data center, but they can also be used to demonstrate
the savings that might accrue from closing a collection of
facilities. Specifically, the Federal Data Center Consolidation
Initiative (FDCCI) anticipates the closure of at least 800 centers
by 2015.[8] Assuming a 500 (computing area)
SF data center and multiplying by our estimated costs, the annual
savings for a single center could range from $140 to $270 thousand,
while the annual savings for all 800 closings could range from a
$112 to $216 million.
Our estimates could be refined with more information about the
individual centers to be closed:[9]
Size of the average center has a direct and
obvious impact on savings. Increasing or decreasing the average
size of a center to be closed could have a proportional affect on
costs. For an actual commercial center (Tier III, unknown computing
intensity) with 16,250 SF of computing area, we confirmed annual
costs of $230 per SF. This is less than the $361 per SF we
estimated for the (much) smaller center, and suggests there could
be scale effects for major costs such as energy, M&R, and
management.
Location determines the labor and utility rates
and climatic conditions that drive many costs. For simplicity, our
examples assume all centers are located in the Washington D.C.
area; In comparison, operations costs in Honolulu were 35 percent
higher than the D.C area in 2010, while costs in Alamogordo, NM
were 28 percent less. Even greater variations can be seen for
international locations.
Level of Service provided substantially affects
operations costs such as M&R and management. Data centers
operated over 80 hours per week create an estimated M&R
requirement roughly 35 percent higher than those operated from 40
to 80 hours, as assumed in the estimates shown in Table 1. With
regard to management costs, we assumed in our example that these
costs are based on standard commercial practice, which is a fixed
percentage of lease costs. If the center is part of a larger
campus, then an in-house staff could handle management
responsibilities for much less (about 80 percent less) than
commercial costs.
In summary, our task was to provide a federal client with an
approximation of the savings that follow from closing a small data
center. We found that these savings can vary broadly by center
design and computing intensity, and we suggested that additional
data on size, location, and service levels could be used to refine
estimates. We also demonstrated that our approach could estimate
aggregate savings for multiple facility closures, such as those
anticipated by the FDCCI.
-Luca Romani
Comment on this post.
To download the full report click here.
[1] Presidential Memorandum-Disposing of Unneeded Federal
Real Estate, June 10, 2010.
http://www.whitehouse.gov/the-press-office/presidential-memorandum-disposing-unneeded-federal-real-estate
[2] Pitt, Seader and Brill, Tier Classifications Define
Site Infrastructure Performance. Uptime Institute, Santa Fe,
NM, 2006.
[3] See
http://www.energystar.gov/index.cfm?c=new_bldg_design.bus_target_finder
[4] MARS is a facility cost forecast system used by federal
agencies and large commercial property holders. See
http://www.whitestoneresearch.com/Mars-Info
[5] The range of computing intensities was derived from
data center benchmarks at http://energybenchmarking.lbl.gov/; and a
presentation by Jonathan Koomey, Data center electricity use:
what we know, EPA Stakeholder Workshop, Santa Clara CA,
February 16, 2007.
[6] We understand that a high computing intensity level in
a 500 SF data center is unlikely but thought that the estimated
cost was instructive.
[7] Our simplest (non tier) center model has a 45 SF
mechanical room.
[8] A recent update on FDCCI activities and a list of
deliverables is found at:
http://www.cio.gov/documents/FDCCI-Update-Memo-07202011.pdf
[9] The alternative costs discussed here are described in
the Whitestone Facility Operations Cost Reference 2010-2011,
Whitestone Research, Santa Barbara CA, August 2010.
Back to
Blog