GreenCloud: A Packet-level Simulator of
Energy-aware Cloud Computing Data Centers
Dzmitry Kliazovich and Pascal Bouvry
FSTC CSC/SnT, University of Luxembourg
6 rue Coudenhove Kalergi, Luxembourg
dzmitry.kliazovich@uni.lu, pascal.bouvry@uni.lu
Yury Audzevich
University of Trento
Sommarive 14, 38100 Trento, Italy
audzevich@disi.unitn.it
Samee Ullah Khan
North Dakota State University
Fargo, ND 58108-6050
samee.khan@ndsu.edu
Abstract—Cloud computing data centers are becoming
increasingly popular for the provisioning of computing resources.
The cost and operating expenses of data centers have skyrocketed
with the increase in computing capacity. Several governmental,
industrial, and academic surveys indicate that the energy utilized
by computing and communication units within a data center
contributes to a considerable slice of the data center operational
costs.
In this paper, we present a simulation environment for
energy-aware cloud computing data centers. Along with the
workload distribution, the simulator is designed to capture
details of the energy consumed by data center components
(servers, switches, and links) as well as packet-level
communication patterns in realistic setups.
The simulation results obtained for two-tier, three-tier, and
three-tier high-speed data center architectures demonstrate the
effectiveness of the simulator in utilizing power management
schema, such as voltage scaling, frequency scaling, and dynamic
shutdown that are applied to the computing and networking
components
1
.
Keywords-energy efficiency, next geneartion networks, cloud
computing simulations, data centers
I. INTRODUCTION
Over the last few years, cloud computing services have
become increasingly popular due to the evolving data centers
and parallel computing paradigms. The notion of a cloud is
typically defined as a pool of computer resources organized to
provide a computing function as a utility. The major IT
companies, such as Microsoft, Google, Amazon, and IBM,
pioneered the field of cloud computing and keep increasing
their offerings in data distribution and computational hosting.
The operation of large geographically distributed data
centers requires considerable amount of energy that accounts
for a large slice of the total operational costs for cloud data
centers. Gartner group estimates energy consumptions to
account for up to 10% of the current data center operational
expenses (OPEX), and this estimate may rise to 50% in the
next few years [1]. However, computing based energy
consumption is not the only power-related portion of the OPEX
bill. High power consumption generates heat and requires an
accompanying cooling system that costs in a range of $2 to $5
million per year for classical data centers.
1
The authors would like to acknowledge the support of Luxembourg
FNR in the framework of GreenIT project (C09/IS/05).
Failure to keep data center temperatures within operational
ranges drastically decreases hardware reliability and may
potentially violate the Service Level Agreement (SLA) with the
customers. A major portion (over 70%) of the heat is generated
by the data center infrastructure; therefore, an optimized
infrastructure installation may play a significant role in the
OPEX reduction.
The first power saving solutions focused on making the
data center hardware components power efficient. Techniques,
such as Dynamic Voltage and Frequency Scaling (DVFS), and
Dynamic Power Management (DPM) [2] were extensively
studied and widely deployed. Because the aforementioned
techniques rely on power-down and power-off methodologies,
the efficiency of these techniques is at best limited. In fact, an
idle server consumes about 2/3 of the peak load [3].
Because the workload of a data center fluctuates on the
weekly (and in some case on hourly basis), it is a common
practice to overprovision computing and communicational
resources to accommodate the peak (or expected maximum)
load. In fact, the average load accounts only for 30% of data
center resources [4]. This allows putting the rest of the 70% of
the resources into a sleep mode for most of the time. However,
achieving the above requires central coordination and energy-
aware workload scheduling techniques. Typical energy-aware
scheduling solutions attempt to: (a) concentrate the workload in
a minimum set of the computing resources and (b) maximizing
the amount of resources that can be put into sleep mode.
Moreover, performing power management dynamically during
runtime considering wide range of system parameters may be
up to 70% more efficient rather than static optimization [14].
Most of the current state-of-the-art research on energy
efficiency has predominantly focused on the optimization of
the processing elements. However, as recorded in earlier
research, more than 30% of the total computing energy is
consumed by the communication links, switching and
aggregation elements. Similar to the case of processing
components, energy consumption of the communication fabric
can be reduced by scaling down the communication speeds and
cutting operational frequency along with the input voltage for
the transceivers and switching elements [5]. However, slowing
the communicational fabric down should be performed
carefully and based on the demands of user applications.
Otherwise, such a procedure may result in a bottleneck, thereby
limiting the overall system performance.
A number of studies demonstrate that often a simple
optimization of the data center architecture and energy-aware
978-1-4244-5637-6/10/$26.00 ©2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
Figure 1. Architecture of the GreenCloud simulation environment.
scheduling of the workloads may lead to significant energy
savings. Ref. [6] demonstrates energy savings of up to 75%
that can be achieved by traffic management and workload
consolidation techniques.
This article presents a simulation environment, termed
GreenCloud, for advanced energy-aware studies of cloud
computing data centers in realistic setups. GreenCloud is
developed as an extension of a packet-level network simulator
Ns2 [7]. Unlike the (only) existing cloud computing simulator
CloudSim [8], GreenCloud extracts, aggregates, and makes
information about the energy consumed by computing and
communication elements of the data center available on an
unprecedented fashion. In particular, a special focus is devoted
to accurately capture communication patterns of currently
deployed and future data center architectures.
The rest of the paper is organized as follows: Section II
presents the main simulator components and related energy
models; Section III focuses on the thorough evaluation of the
developed simulation environment; Sefction IV concludes the
paper providing the guidelines for building energy-efficient
data centers and outlining directions for future work on the
topic.
II. S
IMULATION OF ENERGY-EFFICIENT DATA CENTER
A. Energy Efficiency
From the energy efficiency perspective, a cloud computing
data center can be defined as a pool of computing and
communication resources organized in the way to transform
the received power into computing or data transfer work to
satisfy user demands. Only a part of the energy consumed by
the data center gets delivered to the computing servers directly.
A major portion of the energy is utilized to maintain
interconnection links and network equipment operations. The
rest of the electricity is wasted in the power distribution
system, dissipates as heat energy, and used up by air-
conditioning systems. In light of the above discussion, in
GreenCloud, we distinguish three energy consumption
components: (a) computing energy, (b) communicational
energy, and (c) the energy component related to the physical
infrastructure of a data center.
B. Structure of the Simulator
GreenCloud is an extension to the network simulator Ns2
[7], which we developed for the study of cloud computing
environments. The GreenCloud offers users a detailed fine-
grained modeling of the energy consumed by the elements of
the data center, such as servers, switches, and links. Moreover,
GreenCloud offers a thorough investigation of workload
distributions. Furthermore, a specific focus is devoted on the
packet-level simulations of communications in the data center
infrastructure, which provide the finest-grain control and is not
present in any cloud computing simulation environment.
Fig. 1 presents the structure of the GreenCloud extension
mapped onto the three-tier data center architecture.
C. Simulator components
Servers (S) are the staple of a data center that are
responsible for task execution. In GreenCloud, the server
components implement single core nodes that have a preset on
a processing power limit, associated size of the
memory/storage resources, and contains different task
scheduling mechanisms ranging from the simple round-robin to
the sophisticated DVFS- and DNS-enabled.
The servers are arranged into racks with a Top-of-Rack
(ToR) switch connecting it to the access part of the network.
The power model followed by server components is
dependent on the server state and its CPU utilization. As
reported in [3] an idle server consumes about 66% of its fully
loaded configuration. This is due to the fact that servers must
manage memory modules, disks, I/O resources, and other
peripherals in an acceptable state. Then, the power
consumption increases with the level of CPU load linearly. As
a result, the aforementioned model allows implementation of
power saving in a centralized scheduler that can provision the
consolidation of workloads in a minimum possible amount of
the computing servers.
Another option for power management is Dynamic
Voltage/Frequency Scaling (DVFS) [5], which introduces a
tradeoff between computing performance and the energy
consumed by the server. The DVFS is based on the fact that
switching power in a chip decreases proportionally to V
2*
f.
Moreover, voltage reduction requires frequency downshift.
This implies a cubic relationship from f in the CPU power
consumption. Note that server components, such as bus,
memory, and disks do not depend on the CPU frequency.
Therefore, the power consumption of an average server can be
expressed as follows [11]:
P = P
fixed
+ P
f
*f
3
(1)
Fig. 2 presents the server power consumption model
implemented in GreenCloud. The scheduling depends on the
server load level and operating frequency, and aims at
capturing the effects of both of the DVFS and DPM
techniques.
Switches and Links form the interconnection fabric that
delivers workload to any of the computing servers for
execution in a timely manner.
The interconnection of switches and servers requires
different cabling solutions depending on the supported
bandwidth, physical and quality characteristics of the link. The
quality of signal transmission in a given cable determines a
tradeoff between transmission rate and the link distance, which
978-1-4244-5637-6/10/$26.00 ©2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
Fmax
Fmin
0
0.2
0.4
0.6
0.8
1
Pfixed
Ppeak
CP U F requency
Server load
Power consumption
Fi
g
ure 2. Com
p
utin
g
server
p
ower consum
p
tion.
are the factors defining the cost and energy consumption of the
transceivers.
The twisted pair is the most commonly used medium for
Ethernet networks that allows organizing Gigabit Ethernet
(GE) transmissions for up to 100 meters with the transceiver
power consumed of around 0.4W or 10 GE links for up to 30
meters with the transceiver power of 6W.
The twisted pair cabling is a low cost solution. However,
for the organization of 10 GE links it is common to use optical
multimode fibers. The multimode fibers allow transmissions
for up to 300 meters with the transceiver power of 1W. On the4
other hand the fact that multimode fibers cost almost 50 times
the twisted pair cost motivates the trend to limit the usage of 10
GE links to the core and aggregation networks as spending for
the networking infrastructure may top 10-20% of the overall
data center budget [12].
The number of switches installed depends on the
implemented data center architecture. However, as the
computing servers are usually arranged into racks the most
common switch in a data center is Top-of-Rack (ToR) switch.
The ToR switch is typically placed at the top unit of the rack
unit (1RU) to reduce the amount of cables and the heat
produced. The ToR switches can support either gigabit (GE) or
10 gigabit (10GE) speeds. However, taking into account that
10 GE switches are more expensive and current capacity
limitation of aggregation and core networks gigabit rates are
more common for racks.
Similar to the computing servers early power optimization
proposals for interconnection network were based on DVS
links [5]. The DVS introduced a control element at each port of
the switch that depending on the traffic pattern and current
levels of link utilization could downgrade the transmission rate.
Due to the comparability requirements only few standard link
transmission rates are allowed, such as for GE links 10 Mb/s,
100 Mb/s, and 1Gb/s are the only options.
On the other hand, the power efficiency of DVS links is
limited as only a portion (3-15%) of the consumed power
scales linearly with the link rate. As demonstrated by the
experiments in [13] the energy consumed by a switch and all
its transceivers can be defined as:
∑
=
⋅+⋅+=
R
i
rrportslinecardlinecardschasisswitch
PnPnPP
0
.
, (2)
where P
chassis
is related to the power consumed by the switch
hardware, P
linecard
is the power consumed by any active
network line card, P
r
corresponds to the power consumed by a
port (transceiver) running at the rate r. In Eq. (2), only the last
component appears to be dependent on the link rate while other
components, such as P
chassis
and P
linecard
remain fixed for all the
duration of switch operation. Therefore, P
chassis
and P
linecard
can
be avoided by turning the switch hardware off or putting it into
sleep mode. This fact motivated a combination of DVS scheme
with DNS (dynamic network shutdown) approach.
The proposed GreenCloud simulator implements energy
model of switches and links according to Eq. (2) with the
values of power consumption for different elements taken in
accordance as suggested in [6]. The implemented powers
saving schemes are: (a) DVS only, (b) DNS only, and (c) DVS
with DNS.
Workloads are the objects designed for universal modeling
of various cloud user services, such as social networking,
instant messaging, and content delivery. The execution of each
workload object requires a successful completion of its two
main components: (a) computational and (b) communicational.
The computational component defines the amount
computing resourced required to be provided by the computing
server in MIPS or FLOPS and the duration in time for which
these computing resources should be allocated.
The communicational component of the workload defines
the amount and the size of data transfers that must be
performed before, during, and after the workload execution. It
is composed of three parts: (a) the size of the workload, (b) the
size of internal, and (c) the size of external to the data center
communications.
The size of the workload defines the number of bytes that
being divided into IP packets are required be transmitted from
the core switches to the computing servers before a workload
execution can be initiated.
The size of external communications defines the amount of
data required to be transmitted outside the data center network
at the moment of task completion. However, the internal
communications account for the workloads scheduled at
different servers that have interdependencies. The internal
communications specify the amount of data to be
communicated with a randomly chosen server inside the data
center at the moment of task completion. In fact, internal
communication in the data center can account for as much as
70% of total data transmitted [6].
An efficient and effective methodology to optimize energy
consumption of interdependent workloads is to analyze the
workload communication requirements at the moment of
scheduling and perform a coupled placement of these
interdependent workloads – a co-scheduling approach. The co-
scheduling approach will reduce the number of links/switches
involved into communication patterns.
The workload arrival rate/pattern to the data center can be
configured to follow a predefined distribution, such as
Exponential or Pareto, or can be re-generated from traces log
files. Furthermore, the trace-driven workload generation is
designed to simulate more realistic workload arrival process
capturing also intraday fluctuations [4], which may influence
simulated results greatly.
978-1-4244-5637-6/10/$26.00 ©2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
TABLE I. SIMULATION SETUP PARAMETERS
Parameter
Data center architectures
Two-tier Three-Tier
Three-tier
high-speed
Topologies
Core nodes (C
1
)
Aggregation nodes (C
2
)
Access switches (C
3
)
Servers (S)
Link (C
1
-C
2
)
Link (C
2
-C
3
)
Link (C
3
-S)
16
-
512
1536
10 GE
1 GE
1 GE
8
16
512
1536
10 GE
1 GE
1 GE
2
4
512
1536
100 GE
10 GE
1 GE
Link propagation delay 10 ns
Data center
Data center average load
Task generation time
Task size
Average task size
Simulation time
30%
Exponentially distributed
Exponentially distributed
4500 bytes (3 Ethernet packets)
60 inutes
0 200 400 600 800 1000 1200 1400 1600
0
0.2
0.4
0.6
0.8
1
Server #
Server load
Underloaded servers
DVFS can be applied
Id le serve rs
DNS can be applied
Servers at the
peak load
Figure 3. Server workload distribution with a “green” scheduler.
III. PERFORMANCE EVALUATION
In this section we present case study simulations of an
energy-aware data center for two-tier (2T), three-tier (3T), and
three-tier high-speed (3Ths) architectures.
For comparison reasons we fixed the number of computing
nodes to 1536 for all three topologies while the number and
interconnection of network switches varied. Table I
summarizes the main simulation setup parameters.
In contrast with other architectures 2T data center does not
include aggregation switches. The core switches are connected
to the access network directly using 1 GE links (referred as C
2
-
C
3
) and interconnected between them using 10 GE links
(referred as C
1
-C
2
).
The 3Ths architecture mainly improves the 3T architecture
with providing more bandwidth in the core and aggregation
parts of the network. The bandwidth of the C
1
-C
2
and C
2
-C
3
links in 3Ths architecture is ten times of that in 3T and
corresponds to 100 GE and 10 GE respectively. The
availability of 100 GE links allows keeping the number of core
switches as well as the number of paths in ECMP routing
limited to 2 serving the same amount switches in the access.
The propagation delay of all the links is set to 10 ns.
The task generation events and the size of the tasks are
exponentially distributed with an average task size fixed at
4500 bytes which corresponds to 3 Ethernet packets.
The tasks arrived to the data center are scheduled for
execution using energy-aware “green” scheduler. This “green”
scheduler tends to group the workload on a minimum possible
amount of computing servers. The servers left idle are put into
a sleep mode (DNS) while on the under-loaded servers the
supply voltage is reduced (DVFS).
Fig. 3 presents a workload distribution among severs. The
whole load of the data center (around 30% of its total capacity)
is mapped onto approximately one third of the servers
maintaining load at a peak rate (left part of the chart). This
way, the remaining two thirds of the servers can be shut down
using DNS technique. A tiny portion of the approximately 50
out of 1536 servers which load represents a falling slope of the
chart are under-utilized on average, and DVFS technique can
be applied on them.
Table II presents the power consumption of data center
components. The server peak energy consumption of 301 W is
composed of 130 W
2
(43%) allocated for a peak CPU
consumption and 171 W (56%) consumed by other devices like
memory, disks, peripheral slots, mother board, fan, and power
supply unit [10]. As the only component which scales with the
load is the CPU power, the minimum consumption of an idle
server is bounded and corresponds to 198 W (66%) where also
a portion of CPU power consumption of 27 W required to keep
the operating system running is included.
The switches’ consumption is almost constant for different
transmission rates as most of the power (85-97%) is consumed
by their chassis and line cards and only a small portion (3-
15%) is consumed by their port transceivers. Switch power
consumption values are derived from [6] with a twisted pair
cable connection considered for the rack switch (C
3
) and
2
Chosen based on the specification of Intel Xeon 4-core processor
with 8MB of cache running at 3.33 GHz.
optical multimode fiber for the core (C
1
) and aggregation (C
2
)
switches.
Table III presents simulation results obtained for three
evaluated data center topologies. On average, the data center
consumption is around 503kW·h during an hour of the runtime.
On the yearly basis it corresponds to 4409MW·h or $441k with
an average price of 10c per kW·h.
The processing servers share around 70% of total data
center energy consumption while the communicational links
and switches account for the rest 30% of the total amount.
Furthermore, the consumption of switches breaks with 17%
allocated for core switches, 34% for aggregation switches, and
50% for the access switches. This way, also taking into account
the requirements for network performance, load balancing, and
communication robustness the obvious choice is to keep core
and aggregation switches constantly running at the full speed.
Fig. 4 reports an average distribution of energy consumption in
a data center.
Table IV compares the impact on energy consumption of
DVFS, DNS, and DVFS with DNS schemes applied on both
computing several and networking equipment. The DVFS
scheme alone reduces power consumption to 84% from the
nominal level. Most of the reduction comes from downshifting
CPU voltage as CPU components accounts for 43% of the total
energy consumed by the server. On the other hand DVFS
shows itself ineffective for the switches as only a tiny portion
switch’s energy (3%) is sensitive to the transmission rate
variation.
978-1-4244-5637-6/10/$26.00 ©2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
TABLE II. POWER CONSUMPTION OF DATA CENTER COMPONENTS
Parameter Power consumption (W)
Servers
Server peak 301
Server CPU peak 130
Server other
(memory, peripherial,
mother board, fan,
PSU losses)
171
Server idle 198
Switches
Access
network (C
3
)
Core (C
1
) and Aggregation (C
2
)
Chassis 146 1.5K (10G) 15K (100G)
Linecard - 1K (10G) 12K (100G)
Port tranceiver 0.42 0.3K (10G) 1.6K (100G)
TABLE III. DISTRIBUTION OF DATA CENTER POWER CONSUMPTION
Parameter
Power consumption (kW·h)
Two-tier Three-Tier
Three-tier
high-speed
Data center
Servers
Switches
Core (C
1
)
Aggregation (C
2
)
Access (C
3
)
477.8
351
126.8
51.2
-
75.6
503.4
351
152.4
25.6
51.2
75.6
508.6
351
157.6
56.8
25.2
75.6
TABLE IV. COMPARISON OF ENERGY-EFFICIENT SCHEMES
Parameter
Power consumption (kW·h)
No
energy-
saving
DVFS DNS DVFS+DSS
Data center
Servers
Switches
503.4
351
152.4
486.1 (96%)
340.5 (97%)
145.6 (95%)
186.7 (37%)
138.4 (39%)
48.3 (32%)
179.4 (35%)
132.4 (37%)
47 (31%)
CPU
130W (43%)
Memo ry
36W (12%)
Disks
12W (4%)
Peripherial
50W (17%)
Mother boar
d
25W (8%)
Other
48W (16%)
Computing Servers
301 W
Chassis
36%
Linecards
53%
Port
tranceiver s
11%
Switches
Servers
355kW·h (70%)
Core switches
0.87kW·h (5%)
Aggregation switche s
1.74kW·h (10%)
Acc ess switches
75.6kW·h (15%)
Data center
503kW ·h
Figure 4. Distribution of energy consumption in a data center.
The most effective results are obtained by DNS scheme. It
is equally effective for both servers and switches as the most of
their energy consumed shows no dependency on the operating
frequency. However, in order to utilize DNS scheme
effectively its design should be coupled with the data center
scheduler positioned to unload the maximum number of the
servers.
It should be noted that due to the limited size (4500 bytes)
of the tasks generated by the cloud users the impact of traffic
patterns on the interconnection network became minimized.
This fact in part led to the similarity of energy consumed by
different data center architectures. The effects of variable task
sizes and dense traffic loads in the interconnection network will
be explored in the future works on the topic.
IV. C
ONCLUSIONS
In this paper we presented a simulation environment for
energy-aware cloud computing data centers. Greencloud is
designed to capture details of the energy consumed by data
center components as well as packet-level communication
patterns between them.
The simulation results obtained for two-tier, three-tier, and
three-tier high-speed data center architectures demonstrate
applicability and impact from the application of different
power management schemes like voltage scaling or dynamic
shutdown applied on the computing as well as on the
networking components.
The future work will focus on the simulator extension
adding storage area network techniques and further refinement
of energy models used in the simulated components. On the
scheduling part the analysis will be focused on optimal task
allocation with cooperative scheduling techniques [9].
R
EFERENCES
[1] Gartner Group, available at: http://www.gartner.com/
[2] T. Horvath, T. Abdelzaher, K. Skadron, and Xue Liu, “Dynamic Voltage
Scaling in Multitier Web Servers with End-to-End Delay Control,” IEEE
Transactions on Computers, vol. 56, no. 4, pp. 444 – 458, 2007.
[3] G. Chen, W. He, J. Liu, S. Nath, L. Rigas, L. Xiao, and F. Zhao,
“Energy-aware server provisioning and load dispatching for connection-
intensive internet services,” the 5th USENIX Symposium on Networked
Systems Design and Implementation, Berkeley, CA, USA, 2008.
[4] J. Liu, F. Zhao, X. Liu, and W. He, “Challenges Towards Elastic Power
Management in Internet Data Centers”, Workshop on Cyber-Physical
Systems (WCPS), Montreal, Quebec, Canada, June 2009.
[5] Li Shang, Li-Shiuan Peh, and Niraj K. Jha, “Dynamic Voltage Scaling
with Links for Power Optimization of Interconnection Networks,”
Symposium on High-Performance Computer Architecture, 2003.
[6] P. Mahadevan, P. Sharma, S. Banerjee, and P. Ranganathan, “Energy
Aware Network Operations,” IEEE INFOCOM, pp. 1 – 6, 2009.
[7] The Network Simulator Ns2, avialable at: http://www.isi.edu/nsnam/ns/
[8] R. Buyya, R. Ranjan, and R. N. Calheiros, “Modeling and Simulation of
Scalable Cloud Computing Environments and the CloudSim Toolkit:
Challenges and Opportunities,” 7th High Performance Computing and
Simulation Conference, Leipzig, Germany, June, 2009.
[9] S. U. Khan and I. Ahmad, “A Cooperative Game Theoretical Technique
for Joint Optimization of Energy Consumption and Response Time in
Computational Grids,” IEEE Transactions on Parallel and Distributed
Systems, vol. 21, no. 4, pp. 537-553, 2009.
[10] X. Fan, W.-D. Weber, and L. A. Barroso, “Power provisioning for a
warehouse-sized computer,” ISCA, New York, NY, USA, 2007.
[11] Y. Chen, A. Das, W. Qin, A. Sivasubramaniam, Q. Wang, and N.
Gautam, “Managing server energy and operational costs in hosting
centers,” ACM SIGMETRICS, New York, USA, pp. 303-314, 2005.
[12] A. Greenberg, P. Lahiri, D. A. Maltz, P. Patel, and S. Sengupta,
“Towards a next generation data center architecture: scalability and
commoditization,” ACM Workshop on Programmable Routers For
Extensible Services of Tomorrow, Seattle, WA, USA, August, 2008.
[13] P. Mahadevan, P. Sharma, S. Banerjee, and P. Ranganathan, “A Power
Benchmarking Framework for Network Devices,” IFIP-TC 6
Networking Conference, Aachen, Germany, May, 2009.
[14] B. Khargharia, S. Hariri, F. Szidarovszky, M. Houri, H. El-Rewini, S. U.
Khan, I. Ahmad, and M. S. Yousif, “Autonomic Power and Performance
Management for Large-Scale Data Centers,” IEEE International Parallel
and Distributed Processing Symposium (IPDPS), Long Beach, CA,
USA, March 2007.
978-1-4244-5637-6/10/$26.00 ©2010 IEEE
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE Globecom 2010 proceedings.
