null null

Green Server Design: Beyond Operational Energy to Sustainability
Jichuan Chang, Justin Meza, Parthasarathy Ranganathan, Cullen Bash, Amip Shah
Hewlett Packard Labs
“Green” server and datacenter design requires a focus on
environmental sustainability. Prior studies have focused on
operational energy consumption as a proxy for sustainability,
but this metric only captures part of the environmental
impact. In this paper, we argue that to understand the total
impact, we need to examine the entire lifecycle of the system,
beyond operational energy to also include material use and
manufacturing. We make two main contributions. We present
a methodology that allows such a lifecycle analysis,
specifically providing attribution of sustainability bottlenecks
to individual system architecture components. Using this
methodology, we compare the sustainability tradeoffs
between popular energy-efficiency optimizations and discuss
sustainability bottlenecks and optimizations for future system
designs.
1. Introduction
Environmental sustainability (the manufacturing,
operation, and disposal of products to minimize their
environmental impact in terms of destruction of natural
resources or production of undesired emissions) is fast
becoming an important design constraint for
Information Technology (IT) systems [2]. The carbon
footprint of the IT industry, though only 2% of the
world economy, is estimated to be equal to that of the
entire aviation industry [27]. Even more importantly,
IT is increasingly being used to address the remaining
98% of the carbon emissions of the world economy
[27] (e.g., use of video conferencing to avoid travel)
and as this trend continues, it will become more
important to design ―green‖ IT systems. A recent
estimate showed that up to 75% of organizations will
soon consider sustainability as one of the criteria in
their IT purchases [28]. The UK government is starting
a mandatory Kyoto-style cap-and-trade scheme to curb
energy consumptions of businesses [4] and the US
Congress has similarly been considering various
federal cap-and-trade schemes [3].
There has been a large body of prior work on
reducing the operational electricity consumption of
servers (e.g. [6] [12] [19] [24] [30] [29] [31]). Given
that most of the electricity produced in the world
comes from carbon-intensive sources, these
optimizations can help improve the carbon footprint of
servers and datacenters during operation. However,
these approaches do not address environmental impact
of a system across all the stages of its lifecycle such as
the extraction of raw materials, manufacturing,
transportation, operation, and disposal.
In this paper, we examine the problem of lifecyclebased optimization of future server and datacenter
designs. We make two main contributions – (1) a
methodology to reason about sustainability from a
system architecture perspective and (2) a systematic
analysis of the environmental impact of current designs
across their entire lifecycle and the tradeoffs with stateof-the-art energy-efficiency techniques.
2. Measuring Sustainability: Using
Exergy for Architectural Studies
Numerous schemes exist to quantify the
environmental sustainability of systems. Life-cycle
assessment (LCA), a field that has been in practice for
nearly 50 years [1], involves taking an end-to-end
approach to assessing the environmental impact of a
system across various stages in its lifecycle.
In this paper, we perform lifecycle assessment
using the thermodynamic metric of exergy (available
energy) consumption to reason about sustainability. A
detailed description of exergy is outside the scope of
this paper. However, briefly, unlike energy that is
neither created nor destroyed (1st law of
thermodynamics), exergy is continuously consumed in
the performance of useful work by any real entropygenerating process (2nd law of thermodynamics).
Several previous studies have discussed how this
destruction (or consumption) of exergy is
representative of the irreversibility associated with
various processes [7] [21] and correspondingly, to a
first order, the environmental sustainability [11].
Additionally, models for specific IT systems [18] have
shown that optimizations to reduce lifecycle exergy
consumption often map fairly well to optimizations
based on other types of environmental criteria such as
greenhouse gas emissions, pollution, etc. [32].
Unfortunately,
previous
lifetime
exergy
characterizations
have
estimated
the
total
environmental impact in computer systems based on a
mapping of the system mass or material flows to perunit estimates of the environmental impact burden
[18][34]. Figure 1(a) shows such a breakdown for a
typical server (2-socket Xeon-based server with 4
DIMMs and two 72G HDDs, two 1Gb NICs, and 25%
utilization) using these methods. Such a model is not
very useful for system architects because extending
such a breakdown of exergy to systems architecture
choices is not clear. Since architectural choices may
span multiple stages of the entire system lifecycle,
deciding to use one component over another in a
system will result in (often non-intuitive) changes to
the total system environmental impact due to
differences in the manufacturing process, not just for
the chosen component but also for related components
that interact at the system or datacenter level. An
approach that considers lifecycle exergy consumption
from an architectural perspective is required.
DIMMs, and hard disk drives, as opposed to their
associated raw materials. This enables us to express
the environmental impact of complex ensembles of
diverse sets of materials succinctly in terms of system
architectural choices.
Our approach categorizes exergy1 into three broad
categories – embedded, operational, and infrastructure.
Embedded exergy is the amount of exergy used to
―make‖ a system component. To a first degree, this is
the amount of exergy expended during extraction,
manufacturing, transportation, and recycling. For most
components, the bulk of the embedded exergy is
destroyed during manufacturing as complicated
processes use high quality energy to manufacture
highly-ordered electronic components, and various
chemicals required for making these components
themselves require large amounts of energy to
manufacture. Our model abstracts out the appropriate
exergy destruction values for all of the processes
specific to each component, and then aggregates these
data to discern the overall exergy consumption related
to each architectural component2.
Operational exergy is the amount of exergy spent
during a system‘s operational lifetime. Although the
heat dissipated from the server contains useful work
potential, there are currently no practical techniques to
harness this waste heat and recover this exergy. In this
study, therefore, we assume that operational exergy is
equivalent to the electricity consumed during
operation. To determine operational exergy, for each
component, we use its maximum power rating and
model how its power varies with utilization. We
determined these values from published sources,
internal experiments, and communications with system
designers. This model is similar to that used in other
recent system studies (e.g. [23]) and provides a highorder estimate of the power consumed across different
workloads (varying utilizations). We assume a threeyear lifecycle and 99.99% uptime. Figure 1(c)
summarizes our model parameters.
In most datacenters, the cooling and power
delivery infrastructure accounts for a large fraction of
the total electricity consumption, and consequently, we
account for infrastructure exergy as a separate
category. This takes into account the operational
energy used by CRAC units, chillers, cooling towers
and any other equipment employed in the data center
(a) Process-based breakdown of total exergy
(b) Architecture-based breakdown
Part
Embd. (MJ) Sources
CPU
158
[9] [13] [26]
Chipset
66
[13] [22] [26]
DRAM
726
[13] [33] [26]
PCB
1400
[13] [18] [34]
Chassis
512
[16] [18] [21]
PSU
683
[13] [18]
HDD
546
[13] [18]
Fan
209
[16] [18]
Misc.
420
[20] [34]
Part
Processor
Memory
HDD (15K)
NIC (Gigabit)
Fan
Northbridge
Southbridge
PSU
DC conversion
Misc.
Total
# TDP (W) Idle%
2
95 10%
4
10 50%
2
5 80%
2
6 50%
4
3
0%
1
27.1
0%
1
4.3
0%
1
33 100%
15 100%
10.6 100%
354
-
(c) Sustainability modeling parameters
Figure 1: (a) illustrates previous process-based
approaches to reasoning about sustainability, (b)
illustrates our proposed model to reason about
sustainability based on system architecture
components, (c) summarizes key model parameters.
1
Our work attempts to address these issues by
adopting an architecture-centric approach to measuring
and optimizing the environmental impact of systems.
Specifically, we aggregate raw materials at the
component level, allowing us to evaluate
environmental impact at the granularity of familiar
architectural building blocks such as processors,
More specifically, it is exergy consumption. In this paper, we
loosely use the term exergy to refer to exergy consumption.
2
We aggregate the embedded data from multiple public sources [13]
[34] [33] [20] [9] [22] [16] [26]. Notice these data are derived based
on specific supply chain and component models. Modeling
embedded exergy in a different context should not directly use these
numbers, but rather use the methodology and data sources described
here with new, revised assumptions that are appropriate for the
system being modeled.
2
Workload
Ecommerce 1
Ecommerce 2
Dotcom
Pharmacy
SAP 1
SAP 2
Worldcup 1
Worldcup 2
Consolidation 1
Consolidation 2
Animation farm
(a) Total exergy based exploration
(b) Operational energy based exploration
Utilizations
Mean Peak_Sum
7%
17%
23%
49%
16%
36%
3%
11%
17%
31%
26%
75%
10%
53%
8%
19%
34%
79%
31%
79%
93%
100%
OP (% base)
EP
Con
18% 27%
48% 66%
37% 52%
10% 17%
39% 50%
53% 84%
27% 61%
21% 31%
62% 88%
59% 88%
98% 100%
Total (% base)
EP
Con
36%
25%
57%
63%
49%
49%
31%
16%
51%
46%
61%
82%
42%
60%
38%
28%
68%
87%
66%
86%
98% 100%
(c) Real workloads and efficiencies (winners shaded)
Figure 2: Illustration of tradeoffs between different energy-efficiency optimizations.
infrastructure. (Note that on-board fans are considered
part of server operational power.) We assume that
cooling is provisioned appropriately to handle the
maximum power rating, and we use the widely-used
power usage effectiveness (PUE) metric 3 [14] to
compute
infrastructure
exergy.
The
exergy
consumption related to building the power and cooling
infrastructure in the datacenter is outside the scope of
our model; but, when normalized to a datacenter scale
and across multiple IT refresh cycles, we expect the
allocation of its embedded burden is minimal.
categories of optimizations: (i) Energy proportionality
(EP) [6] in the datacenter space has gained a lot of
attention with several optimizations [15] [8] [29] [24]
that seek to make the energy consumed by a system be
proportional to the activity in the system. (ii)
Consolidation (Con) is another optimization common
in current datacenters. The intuition is that typical
utilization on many enterprise services is relatively low
and bursty and that across a collection of systems,
peaks are often unsynchronized (the peak of the sums
of the individual utilizations is lower than the sum of
the peak individual utilizations). Multiple virtual
machines (or tasks in a task scheduler) on separate
servers can be consolidated onto a single server, raising
its utilization and reducing the required server count
(and total power) [25] [30]. (iii) Recently, there have
been several low-power server solutions (LP) based on
energy-efficient, but lower-power processors [23] [17]
[10] [5]. A common idea behind these solutions is to
better match the processor architecture to the workload
characteristics (primarily around CPU-I/O balance) to
leverage significantly better performance/watt.
Figure 2 shows our results from examining these
three optimizations for a parameterized design space
exploration. The benefits from EP are primarily a
function of workload average utilization. Figure 2(a)
shows this design space exploration for an average
workload utilization held constant at 25%. (We
examine other utilization points as well, but omit them
for brevity.) For a given average workload utilization,
we identify different tradeoffs for the LP designs by
using a performance/watt multiplier on the X axis. For
some workloads (e.g, enterprise workloads), a lowerpower processor may lose more in performance than it
saves in power; for these cases the performance/watt
multiplier is less than 1 (right side of the figure),
indicating the LP solution‘s performance/watt (or
energy efficiency) at peak load is worse than a
conventional server. For web workloads, prior studies
[23] [17] [10] have found LP to yield better multipliers
ranging from 2 to 5 (left side of the axis). As discussed
earlier, the effectiveness of consolidation is a function
of how many processes can be packed into a single
server, which in turn is a function of the peak-of-sum
utilization specific to the workload. The Y axis shows
3. Evaluating the state-of-the-art
Exergy breakdown
Figure 1(b) shows the breakdown of total lifecycle
exergy using our models. We focus on the same server
as in Figure 1(a), and assume a workload utilization of
25%, and a PUE (1.6) based on prior studies [17]. The
results show that operational exergy dominates the total
exergy of the system (53%), followed by infrastructure
exergy (27%), and embedded exergy (20%). Of note is
that the embedded exergy contributes a sizable amount
to total system exergy. The dominant components of
embedded exergy are from silicon-based processes and
PCB design. Assuming a datacenter container with
1056 of these servers, the total amount of exergy
consumption is 25.4 Tera Joules over a three year
timeframe, equivalent to approximately 870 metric tons
of coal consumption.
Design space exploration
There has been a large body of prior techniques
that address operational energy. However, their impact
on total exergy hasn‘t been studied. Specifically, how
do these techniques compare from a sustainability point
of view? Are there tradeoffs between operational
exergy and embedded exergy that make some of these
techniques less effective in improving net
sustainability? If these techniques are aggressively
applied to future systems, what would the new
breakdown of exergy consumption look like? To
answer these questions, we studied three broad
3
PUE = 1+ infrastructure_power / operational_power
3
this parameter. Lower values indicate that the peaks are
completely non-synchronized and consolidation can
more readily be leveraged. Different points on the heat
map thus represent different workload/system
configurations.
For each data point, we individually compute the
total exergy for EP, Con, and LP designs providing the
same aggregate performance and identify the
optimization that achieves the best exergy. (Recall that
lower exergy consumption is better.) The heat map‘s
color gradation reflects the absolute value of this best
exergy. The division of the heat map into various
regimes shows the technique that achieves the best
exergy for that region of workload/system
configurations. For energy proportionality, we studied
a best-case future model where all the hardware shows
ideal proportionality (the power consumed in an idle
state is zero). For consolidation, we assumed perfect
bin-packing that minimizes the number of servers.
Figure 2(b) shows a similar picture, but for a case
where only operational energy is considered. For EP
and Con, we model component power after a
conventional server shown in Figure 1(c); for LP, we
model an HP BC2500 blade server with maximum
component powers similar to [23]. Here we assume a
PUE of 1.5 for infrastructure exergy, and adjust the
embedded exergy consumption values of components
within each system based on a scaling of key physical
attributes for each component4.
Observations
This way of representing the data reveals several
interesting high-level trends. First, the figures
individually show the different regions when different
techniques work best and the cross-over points, as well
as the relative magnitude of the benefits. Comparing
the two figures allows us to examine the changes to
these design tradeoffs when optimizing for just
operational energy versus considering total exergy.
Figure 2(a) shows that in general the total exergy
of the system is minimized when going towards the
bottom left region of the graph—not surprising
considering this assumes more power-efficient
components and lower resource activity (more
consolidation). First comparing EP and Con, we
observe that EP outperforms Con when the workloads
are not bursty and don‘t lend themselves to packing
(top right part of Figure 2(a)). The break-even point is
roughly corresponding to workloads with peak-of-sum
utilizations close to 50%. Below this, Con is a better
design alternative. Interestingly, this conclusion is
different than when just focusing on operational
energy. There, given fragmentation in bin-packing,
perfect energy proportionality is always better than
consolidation. However, when considering total
exergy, a reduction in materials associated with fewer
servers provides additional reductions in embedded
exergy that allow Con to be better than EP 5.
Comparing with LP, we find that after a breakeven
point roughly corresponding to 1.6-2.6X improvement
in performance/watt, LP designs are always better than
both EP and Con. Considering the differences between
Figures 2(a) and Figure 2(b), the inflection point at
which LP is better than other alternatives shifts to the
left (requiring even more energy efficiency from lower
power processors) when total exergy is considered.
This is because of the increased embedded exergy from
the larger number of lower-power servers required for
the same performance. Comparing LP and Con, it is
worth noting that there is now a region where
consolidation of multiple small processes into one
server is better than distributing them into multiple
small low-power blades.
Notice that because LP and EP are independent of
peak of sum utilization, the break-even point between
these solutions is dictated entirely by the
performance/watt multiplier. This implies that the
optimal choice between these two solutions is
dependent on their relative energy efficiencies for the
type of workload. The number of machines used in
Con depends on the peak-of-sum utilization, but notice
that consolidation also raises overall system utilization,
increasing operational exergy. This trade-off between
fewer machines and higher utilization is shown as the
angled line dividing LP and Con.
The table in Figure 2(c) illustrates the tradeoffs
between EP and Con with data from various real-world
traces. (They correspond to specific real-world points
in the bottom right portions of the heat maps.) From an
operational exergy perspective, EP achieves more
savings compared to Con for all the enterprise traces,
but, by contrast, from a total exergy perspective, in
many cases Con outperforms EP.
4. Discussion
The results above illustrate that focusing on the
most efficient system design for operational energy
does not always produce the most sustainable solution.
Tradeoffs between operational exergy and embedded
exergy need to be considered. The examples in the
previous section—requiring larger factors of energy
4
For example, we find that the key physical attribute governing the
footprint of a microprocessor is the area of the silicon. Thus, we
normalize the impact calculated in Fig. 1(c) by the area to derive an
‗impact factor‘ representing the exergy consumption per unit area.
This impact factor can then be scaled as required for processors of
different sizes, assuming uniform thickness; fabrication; etc. If other
key attributes vary (e.g., a change in the thickness of the package),
these can be accordingly parameterized as well. A similar approach
can be repeated for each of the different architectural components.
5
Note that we assume a model where consolidation leads to lower
provisioning of servers; if consolidation just allowed servers to be
turned off, we would not get the embedded savings.
4
[2] Revolutionizing Datacenter Energy Efficiency. McKinsey, 2008.
[3] Regional Greenhouse Gas Initiative. http://www.rggi.org. 2008.
[4] UK Government. Carbon Reduction Commitment, July 2009.
[5] D. Andersen, J. Franklin, et al. FAWN: a fast array of wimpy
nodes. SOSP 2009.
[6] L. A. Barroso and U. Hölzle. The case for energy-proportional
computing. IEEE Computer, 40(12):33–37, 2007.
[7] A. Bejan. Advanced Engineering Thermodynamics (2nd
Edition). John Wiley & Sons, 1997.
[8] R. Bianchini and R. Rajamony. Power and energy management
for server systems. IEEE Computer, 37(11):68–74, 2004.
[9] S. Boyd, A. Horvath, et al. Life-cycle energy demand and global
warming potential of computational logic. Env. Sci. Tech., 2009.
[10] A. Caulfield, L. Grupp, and S. Swanson. Gordon: using flash
memory to build fast, power-efficient clusters for data-intensive
applications. ASPLOS-XIV, 2009.
[11] I. Dincer and M. Rosen. Exergy: Energy, Environment and
Sustainable Development. Elsevier, 2007.
[12] X. Fan, W-D. Weber, and L. Barroso. Power provisioning for a
warehouse-sized computer. ISCA 2007.
[13] R. Frischknecht, et al. The ecoinvent database: overview and
methodological framework. J. Life Cycle Assessment, 10(1), 2005.
[14] The Green Grid. Green grid metrics: Describing datacenter
power efficiency. http://www.thegreengrid.org, 2007.
[15] M. Gupta, S.Singh. Greening of the Internet. SIGCOMM 03.
[16] T. Gutowski, et al. Thermodynamic analysis of resources used
in manufacturing processes. Env. Sci. Tech., 43(5), 2009.
[17] J. Hamilton. Cooperative expendable micro-slice servers: Low
cost, low power servers for internet-scale services. CIDR 2009.
[18] C. Hannemann, et al. Lifetime exergy consumption as a
sustainability metric for enterprise servers. ASME ICES, 2008.
[19] T. Heath, et al. Mercury and Freon: Temperature emulation and
management for server systems. ASPLOS-XII 2006.
[20] Y. Huang, C. Weber, and H. Matthews. Carbon footprinting
upstream supply chain for electronics manufacturing and computer
services. IEEE ISSST 2009.
[21] D. Morris, J. Szargut and F. Steward. Exergy analysis of
thermal, chemical and metallurgical processes. Hemisphere, 1988.
[22] N. Krishnan, et al. A hybrid life cycle inventory of nano-scale
semiconductor manufacturing. Env. Sci. Tech., 42(8), 2008.
[23] K. Lim, P. Ranganathan, et al. Understanding and designing
new server architectures for emerging warehouse-computing
environments. ISCA 2008.
[24] D. Meisner, B. Gold, and T. Wenisch. PowerNap: Eliminating
server idle power. ASPLOS-XIV, 2009.
[25] R. Nathuji and K. Schwan. VirtualPower: coordinated power
management in virtualized enterprise systems. SOSP 2007.
[26] J. Oliver, R. Amirtharajah, et al. Life cycle aware computing:
Reusing Silicon Technology. IEEE Computer, 40(12), 2007.
[27] The Climate Group and GeSI, SMART2020: Enabling the low
carbon economy in the information age, 2008.
[28] D. Plummer, et al., Gartner‘s top predictions for IT
organizations and users: Going green and self-healing, 2008.
[29] R. Raghavendra, et al. No ―power‖ struggles: Coordinated
multi-level power management for the data center. ASPLOS 2008.
[30] K. Rajamani and C. Lefurgy. On evaluating request-distribution
schemes for saving energy in server clusters. ISPASS 2003.
[31] P. Ranganathan, P. Leech, D. Irwin, and J. Chase. Ensemblelevel power management for dense blade servers. ISCA 2006.
[32] A. J. Shah, C. D. Patel, , and V. P. Carey. Exergy-based metrics
for sustainable design. IEEES-4, 2009.
[33] E. Williams. The environmental impacts of semiconductor
fabrication. Thin Solid Films, 461(1), 2004.
[34] E. Williams. Energy intensity of computer manufacturing:
Hybrid assessment combining process and economic input-output
methods. Env. Sci.Tech., 38(22), 2004.
efficiency improvement for low-power servers to be
sustainably better, or consolidation being more
sustainable than energy proportionality—illustrate this
point. The best way to optimize for sustainability is to
use power-efficient and material-efficient systems that
scale power with resource usage and are utilized fully.
In future systems, as the ratio of embedded exergy
to total exergy grows, new optimizations will be needed
that explicitly target embedded exergy. For example,
upcycling (reusing components when they would
normally be recycled or discarded) is an effective way
to reduce embedded exergy, amortizing the destruction
of exergy over a longer period of time. However, this
will require new ways of building systems, including
designs that allow technology upgrades to be localized
only to the components that need to be upgraded,
allowing the rest to be upcycled. ―Dematerialization‖
techniques that reduce the material in the solution will
also be important. This will require identifying the
sweet spot of resources for best performance
efficiency.
For
example,
smaller
memory
configurations could use less silicon and consequently
reduce the embedded exergy associated with memory.
Finally, when considering the approaches above, it
is important to note that embedded exergy, operational
exergy, infrastructure exergy, and performance are not
independent variables. For example, dematerialization
sometimes reduces infrastructure exergy consumption
(e.g., removal of sheet metal in the backplane can
enable better designed air flow), but in other cases
increases infrastructure exergy (e.g., removal of fans in
a server can increase overall cooling energy in the
datacenter). Similarly, different optimizations can have
different tradeoffs on performance: backplane redesign
for dematerialization can impact networking
topologies, reductions to cooling infrastructure may
lead to performance throttling, and so on. It will
therefore be important to address sustainability
holistically across the various components of total
lifecycle exergy.
Overall, as sustainability becomes a more important
design consideration for future systems, design
methodologies and system optimizations need to
correspondingly change to address these emerging
challenges. This paper takes the first steps in this
direction—around a methodology to reason about
sustainability bottlenecks from an architectural
viewpoint, and enabling an understanding of tradeoffs
and bottlenecks in future designs. We believe,
however, that we have only scratched the surface and
that these areas offer a rich opportunity for more
innovation by the broader community.
References
[1] ISO 14040: Environmental management – Life Cycle
Assessment – Principles and framework. ISO, 2006
5

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project