Energy Aware Clouds Scheduling Using Anti-load Balancing
Algorithm - EACAB
Cheikhou Thiam, Georges Da-Costa and Jean-Marc Pierson
Institut de Recherche en Informatique de Toulouse, Universite de Toulouse 3 Paul Sabatier,
118 Route de Narbonne, Toulouse, France
Keywords:
Energy, Virtual Machines, Cloud, Consolidation.
Abstract:
Cloud computing is a highly scalable and cost-effective infrastructure for running HPC, enterprise and Web
applications. However rapid growth of the demand for computational power by scientific, business and web-
applications has led to the creation of large-scale data centers consuming enormous amounts of electrical
power. Hence, energy-efficient solutions are required to minimize their energy consumption. The objective of
our approach is to reduce data center’s total energy consumption by controlling cloud applications’ overall re-
source usage while guarantying service level agreement. This article presents Energy aware clouds scheduling
using anti-load balancing algorithm (EACAB). The proposed algorithm works by associating a credit value
with each node. The credit of a node depends on its affinity to its jobs, its current workload and its com-
munication behavior. Energy savings are achieved by continuous consolidation of VMs according to current
utilization of resources, virtual network topologies established between VMs and thermal state of computing
nodes. The experiment results show that the cloud application energy consumption and energy efficiency is
being improved effectively.
1 INTRODUCTION
Up to now, the problem of efficiently allocating
tasks in clusters has received considerable atten-
tion. Task scheduling algorithms have been pro-
posed to optimize the placement of tasks with re-
spect to performance-related criteria. Usually, those
researches do not use migration to consolidate work-
load.
In recent years, many authors have studied the
problem of power aware placement, finding theoret-
ical solutions as well as practical ones(Lawson and
Smirni, 2005)
In a similar field of research, performance of dis-
tributed systems, load-balancing techniques are often
used in order to guaranty good performance. These
techniques work by spreading load on all available
servers, which is efficient from the performance point
of view, but not from the energy point of view.
Typical performance measures include task re-
sponse time, throughput and processor utilization.
But when the goal is to reduce energy consumption,
this type of algorithms can lead to have hosts largely
under-loaded and therefore consuming energy unnec-
essarily.
In this context, this article proposes an energy-
aware anti-load balancing algorithm. This run-time
algorithm will migrate tasks while they are running,
using on-the-fly migration technology.
Classical anti-load balancing algorithms based on
migration techniques(Thiam and Da Costa, 2011) use
mainly one single parameter: Threshold. Depending
on a server load, if over a threshold, load is migrated
on other servers or a new server is switched-on in
order to keep a good quality of service. If under a
threshold, load is migrated on other servers and the
host is switched off.
Most researches on energy efficiency try to reduce
energy consumption of servers, but they usually do
not take into account the cost of cooling systems and
related infrastructure. Thus it is important to set a
minimum threshold to consolidate tasks, but also to
avoid to load hosts heavily by concentrating too many
tasks on the same host.
The main contribution of this work is to propose
an efficient algorithm that migrate tasks to reduce
energy-consumption while preserving performance
and preventing hot spots. The algorithm, called EA-
CAB, is based on results from the field of load balanc-
ing, especially on ALB (Thiam and Da Costa, 2011)
82
Thiam C., Da-Costa G. and Pierson J..
Energy Aware Clouds Scheduling Using Anti-load Balancing Algorithm - EACAB.
DOI: 10.5220/0004856600820089
In Proceedings of the 3rd International Conference on Smart Grids and Green IT Systems (SMARTGREENS-2014), pages 82-89
ISBN: 978-989-758-025-3
Copyright
c
2014 SCITEPRESS (Science and Technology Publications, Lda.)
algorithm. We evaluate EACAB by simulation using
Enersim.
The rest of the paper is organized as follows. Sec-
tion 2 discusses related work, followed by the model
in Section 3. The proposed migration algorithms are
discussed in Section 4. An analysis of simulation re-
sults of the proposed algorithms is presented in Sec-
tion 5.
2 RELATED WORK
In this section, we present energy consumption mod-
els and migration techniques.
2.1 About Energy Consumption
In the past few years, people start to realize that the
energy consumption is a critical issue since energy
demand of clusters have been steadily growing fol-
lowing the increasing number of data centers. Several
strategies for energy saving in heterogeneous clus-
ters have been proposed and studied. Recently, many
energy-aware scheduling algorithms have been devel-
oped primarily using the dynamic voltage-frequency
scaling (DVFS) capability which has been incorpo-
rated into recent commodity processors. However,
these techniques are rarely compatible with optimiz-
ing both quality of service for tasks and energy con-
sumption.
Chase et al. (Chase et al., 2001) illustrated a
method of determining the aggregate system load and
the minimal set of servers that can process a load.
A similar idea leverages work in cluster load balanc-
ing to determine when to turn machines on or off to
handle a given load (Pinheiro et al., 2001). A crit-
ical problem for these ideas is that in order to turn
lightly loaded machines off or to assign workload to
newly turned-on machines, the task need to be trans-
ferred from one machine to another. But almost all
the operating systems used in the real clusters, e.g.
Windows, Unix and Linux, cannot support such kind
of operations. So in their research specific OS fea-
tures have to be developed and applied, which in turn
limits the practicability of their approaches. Research
efforts are in great need to architect green data cen-
ters with better energy efficiency. The most promi-
nent approach is the consolidation enabled by virtual-
ization. Server virtualization technology has become
known to improve utilization while reducing various
aspects of consumption and offering the ability to ride
the green trend while helping businesses save money.
There are indeed a few high-performance computers
designed with energy-saving in mind, such as Blue-
Gene/L (Adiga et al., 2002), which uses a system
on a chip to reduce energy consumption, and Green
Destiny (Warren et al., 2002), which uses low-power
Transmeta nodes. But their concern on energy sav-
ing is only confined to the design of hardware, with
nothing to do with the strategies for power control at
run-time, which also plays an important role. There
is also a large effort in saving energy for desktop and
mobile systems. In fact, most of the early researches
in energy-aware computing were on these systems.
At the system level, there has been work in trying to
make the OS energy-aware by making energy the first
class resource (Ellis, 1999). A number of good meth-
ods and ideas in these studies could be introduced to
the energy saving schemes in cluster systems.
2.2 Job Scheduling
Migration allows tasks to be moved from their orig-
inally assigned hosts to another one, at runtime.
Several virtualization software support virtual ma-
chine migration and switching off unused hosts to re-
duce energy consumption. Entropy(Hermenier et al.,
2009), a Consolidation Manager for Clusters, is a re-
source manager for homogeneous clusters, which per-
forms dynamic consolidation of resources based on
constraint programming, using VM migration. (Be-
loglazov and Buyya, 2010) proposes a novel tech-
nique for dynamic consolidation of VMs based on
adaptive utilization thresholds which reduces Service
Level Agreements (SLA) violation. Also, vendors,
like VMware with vSphere 4, includes a Distributed
Power Management (DPM) that monitors virtual ma-
chines resource utilization within the cluster.
In (Srikantaiah et al., 2008) the proposed algo-
rithm aims at finding a minimal energy allocation of
workload to servers. In all these studies the objective
is to minimize the energy consumption of the servers,
while satisfying given performance-related bounds on
the period. In (Pierson and Casanova, 2011), a model
of cluster hosts that have DVFS capacities was used
to calculate a bound on the optimal solution. Contrary
to our work, those researches only minimize servers
energy consumption while not taking into account the
impact of hot spots on the cooling system. Proposed
work considers the algorithm of load unbalancing to
improve tasks management, taking into account the
cost of migrations.
EnergyAwareCloudsSchedulingUsingAnti-loadBalancingAlgorithm-EACAB
83
3 MODEL AND OBJECTIVES
In this section we consider a group of clusters that
consists of H computing hosts, or hosts. We model
each host with the unit: CPU cycles per time-unit. We
also consider T services, or tasks, that run on the clus-
ter. A task l is defined by its percentage of CPU need,
l. The CPU need of a task is the CPU share it would
use on a host that is dedicated to it. We will define
multiple objective functions. They define optimiza-
tion problems taking into account constraints that will
result in efficient load unbalancing algorithms. We
use an algorithm which is the opposite of load balanc-
ing because our main objective is to increase energy
gain.
3.1 Model and Hypothesis
We consider an environment represented by a large-
scale data center consisting of H =
N
i=1
H
i
hetero-
geneous physical hosts. Each cluster i has H
i
hosts.
There are N clusters. Each host is characterized by
the CPU performance defined in Millions Instructions
Per Second (MIPS). We consider T
i
tasks, that run on
the cluster i.
Thus, migration must take place under several
constraints:
Conservation of the execution context: It must be
possible to stop the execution process of the task
and restart it where it has stopped. We must be
able to get the task execution context (MIPS, size,
remaining size, state, memory, etc.), to transfer
this state via the network, to reload and restart the
task. The migration of virtual machines is used to
obtain this result;
Slowdown prevention : Job slowdown increases
the execution time and therefore increases the en-
ergy consumed by hosts and impact users;
Avoiding overloaded hosts to avoid heat points as
it increases energy consumption of cooling. To
measure the saturation of a cluster, we use the
over-loaded threshold ε, which we call satura-
tion. A cluster reaches saturation when its load
is greater than ε.
The proposed algorithm provides solutions to
these problems. The following steps are executed by
the scheduler:
Estimate the requested load (R
i
) of the cluster i.
This load depends on the number of task (T
i
) exe-
cuted by the cluster and their load (l
i, j,k
is the re-
quested load of task k in cluster i on host j). r
i, j
is
the aggregated load of all tasks on host j in cluster
i.
R
i
=
H
i
j=1
(r
i, j
) r
i, j
=
T
i
k=1
(l
i, j,k
)
If task k is not running on host j in cluster i, then
l
i, j,k
=0.
Computes load C
i
and speed V
i
of cluster i
C
i
=
H
i
j=1
(c
i, j
)
V
i, j
=
H
i
j=1
(v
i, j
)
c
i, j
Actual load of host j in cluster i
v
i, j
Maximum speed of host j in cluster i in Mips
Job satisfaction S
i
of cluster i (same for task
satisfaction of host j in cluster i)
S
i
=
C
i
R
i
constraints
j, k l
i, j,k
[0, 1],
i, j r
i, j
[0, S
i
]
i S
i
]0, 1]
hypotheses
Communication within a cluster and between
other clusters are considered as negligible;
For each cluster there is at least one host and
one task;
Migration cost is considered as the same in
CloudSim. To migrate a VM, only RAM has
to be copied to another node. The migration
time depends on the size of RAM and the avail-
able network bandwidth. M migration delay =
RAM / bandwidth + C (C = 10 sec). Bandwidth
is considered as constant;
To minimize energy consumption the load of
each host tries to verify : i, j c
i, j
{0}
[γ, 1]
γ is the underloaded threshold
This equation means that each host are either
switched off or with a load over γ;
Besides the under-load threshold we add an-
other parameter ε corresponding to saturation
and acting as a overload threshold. To verify
that there is no over-loaded hosts :
i, j c
i, j
ε
We will use the classical linear model of power
consumption in function of load :
i, j P
i, j
= P
i, j
min
+ c
i, j
(P
i, j
max
P
i, j
min
) Therefore the
total power consumption of the system is: P =
N
i=1
H
i
j=1
P
i, j
To obtain energy consumed during a
time slice, instantaneous power has to be multiplied
by time. Total energy is then obtained by summing
SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems
84
all the energy of those time slices. An objective is to
minimize this energy.
3.2 Objectives
The main objective of our approach is to improve
cloud’s total energy efficiency by controlling cloud
applications’ overall energy consumption while en-
suring cloud applications service level agreement.
Therefore, our work must take place under several ob-
jectives :
Ease of task Management : Managing jobs can
be a tremendous task which requires many highly
experienced IT experts. Providing an easily con-
figurable system can significantly reduce the costs
and ease the system management. One of our
goals is to design a system which requires mini-
mal human intervention to be configured. More-
over, once the system is deployed and configured,
it becomes increasingly important to perform up-
dates and/or add new servers. In such scenar-
ios servers will be required to brought offline and
added back later. Our goal is to design a system
which is flexible enough to allow for dynamic ad-
dition and removal of servers. Finally, as system
components can fail at any time, it is desirable for
a system to heal in the event of failures without
human intervention. Consequently, we aim at de-
signing a system using self-healing mechanisms
to enable high availability.
Energy Efficiency: Over the past years, rising en-
ergy bills have resulted in energy efficiency to
become a major design constraint for distributed
systems providers. Given that traditional clouds
are rarely fully utilized, significantly energy sav-
ings can be achieved during periods of low uti-
lization by transitioning idle servers in a power
saving state. However, as servers are rarely fully
idle, first idle times need to be created. One of
our goals is to propose task placements manage-
ment algorithms which are capable of creating
idle times, transitioning idle servers in a power
saving state and waking them up once required
(e.g. when load increases).
Avoiding overloaded hosts to avoid heat points
as it increases energy consumption of cooling.
To measure the saturation of a cluster, we use
the over-loaded threshold ε, which we call sat-
uration. A cluster reaches saturation when its
load is greater than ε.
4 ENERGY AWARE CLOUDS
SCHEDULING USING
ANTI-LOAD BALANCING
ALGORITHM (EACAB)
Selection policy selects appropriate tasks for migra-
tion. Location policy determines then suitable hosts to
receive them. In other words, they locate complemen-
tary hosts to/from which they can send/receive tasks.
Current version of our algorithm uses tasks load to
take those decisions. In the following, we present an
algorithm (EACAB) based on the merged principle of
Comet(Jeon et al., 2010) and of anti-load balancing.
4.1 Credit based Anti-load Balancing
Model
The algorithm proposed in this article aims at maxi-
mizing Credit which is a value used when calculating
the energy-efficiency of the system behavior.
This Credits algorithm is an adaptation of the The
Comet Algorithm (Chow and Kwok, 2002), a load bal-
ancing algorithm. Comet is based calculating credit
for mobile agent. Each agent in Comet is trying to
maximize its own credit by moving between hosts.
An agent a
i
uses the following formula:
C
i
= x
1
w
i
+ x
2
h
i
x
3
g
i
Where w
i
: computation load of the host running
agent a
i
, h
i
and g
i
: communication load inside and
outside agent a
i
, and where x
1
, x
2
and x
3
are posi-
tive float coefficients which constitute dependence as-
signed to each agent from its creation to estimate its
affinity relative to their host. Thus an agent will move
to a new host if it result in a lower host load, or if it
reduces external communication or if it increases in-
ternal communication. This algorithm does not take
int account migration cost.
In the same way, the proposed algorithm in this
article works by associating a credit value with each
host. The credit of a host depends on the host, its cur-
rent workload, its communications behavior and his-
tory of task execution. When a host is under-loaded
(load < globally defined threshold), all its tasks are
migrated to a comparatively more loaded host.
In dynamic load unbalancing schemes, the two
most important policies are selection policy and lo-
cation policy. Selection policy concerns the choice of
the host to unload. Location policy chose the desti-
nation host of these moved tasks. An important char-
acteristic of selection policy is to prevent the desti-
nation host to become overloaded. Also, migration
costs must be compensated by the performance im-
provement.
EnergyAwareCloudsSchedulingUsingAnti-loadBalancingAlgorithm-EACAB
85
Each host has its own Credit, which is a float
value. The higher a host Credits, the higher its chance
its tasks to stay at the same host. The credit of a host
increases if:
Its workload or the number of tasks in the host
increases;
Communication between its tasks and other hosts
increases;
Its load increases while staying between the
under-loaded threshold γ and the over-loaded
threshold ε.
On the contrary, the credit of a host decreases in the
cases below:
Its workload or its number of tasks decrease;
It has just sent or received a message from the
scheduler which indicates that the host will prob-
ably become empty in a short while.
The Credit of a host will be used in the selection pol-
icy: the host which credit is the lower is selected for
tasks migration. The location policy identifies the re-
mote host with the highest credit which is able to re-
ceive the tasks selected by the selection policy with-
out being over-loaded. Figure 1 shows an example of
migration. The percentage represents here the occu-
pancy rate of each task on the host.
Figure 1: Different contexts for a migration.
4.2 Algorithm Description
In Comet mobile agents move between hosts accord-
ing to their affinities (credit) to achieve load balanc-
ing. Here we work with tasks which migrate depend-
ing on the load of the host. We apply the Credit con-
cept to the migration of tasks. EACAB algorithm
is based on the technique of calculating credit (σ
i, j
)
of each host ( j in cluster i) by the same method of
Comet(Jeon et al., 2010). In EACAB, the formula is
then:
σ
i, j
= c
i, j
r
i, j
t
i, j
+ ε γ
Were c
i, j
is the actual load of host j in cluster i,
r
i, j
is its requested computation load, t
i, j
is its task
satisfaction, and γ and ε are respectively under-load
and over-load threshold.
EACAB provides task scheduling strategy, which
dynamically migrate tasks among computing hosts,
transferring tasks from underloaded hostssec4 to
loaded but not overloaded hosts. It balances load of
computing hosts as far as possible in order to reduce
program running time.
The decision making algorithm behaves globally
as follows:
If σ
i, j
< 0 , the host j of cluster i is over-loaded or
under-loaded.
If c
i, j
> ε , the host j of cluster i is over-loaded
If c
i, j
< γ , the host j of cluster i is under-loaded
This algorithm is described in Algorithm 1. For
the sake of simplicity, corner cases such as all nodes
over-loaded are not included. Selection policies take
into account credits and migration cost. The selected
host (node j
in cluster i
) is the one with the min-
imun σ
i
, j
weighed by the migration cost between the
current position of the job and the potential host. If
τ
i, j,i
, j
is the migration cost between the node j in
cluster i and the node j
in cluster i
, the selected host
is the one that minimize : σ
i
, j
.
τ
i, j,i
, j
Max
i
′′
, j
′′
(τ
i, j,i
′′
, j
′′
)
4.3 Analysis of the Algorithm
This migration algorithm’s goal is to minimize the en-
ergy. It is composed of two parts. In the first part, it
checks for each host i if the load is below the thresh-
old. If this is the case, it locates the host j that will
receive all tasks of host i. The second part manages
hotspots. To reduce the load of an overloaded host, it
begins to migrate the slowest task. Selection policy
will choose the task that will stay the longest on the
host. Policy of localization will then identify the host
that will receive the task without exceeding its capac-
ities (ie. its load after migration will still be under ε).
So this host will be the new destination of the task.
5 EXPERIMENTS AND RESULTS
In order to evaluate the gains of EACAB compared
to classical algorithms, we implemented this algo-
rithm in EnerSim. This simulator is based on Grid-
Sim(Calheiros et al., 2011), where we added power
consumption (extended from CloudSim(Buyya and
Murshed, 2002) implementation) and virtual machine
(mainly their migration). It is a java event driven sim-
ulator of grids, cluster and Clouds. It provides infor-
mation about execution times, but also about instan-
SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems
86
Algorithm 1: Energy aware clouds scheduling using
anti-load balancing algorithm (EACAB).
Calculate c
i, j
, σ
i, j
// Load, credit of node j in
cluster i
Sort in ascending order hosts according to the value
of their credit
for (M
orig
in sorted host j in all cluster i) do
Update
c
i, j
,
σ
i, j
for remaining hosts
Sort the remaining hosts (per load)
if (c
i, j
< γ) then
// In case M
orig
is under-loaded
for (M
dest
in all host j
in all cluster i
) do
if M
orig
̸= M
dest
and (c
i
, j
γ) and (c
i, j
+
c
i
, j
< ε) then
Add M
dest
to potential destination set
Potential
end if
Migrate all task from M
orig
to the element
with lower credits weighted by migration
cost in Potential
end for
else
// M
orig
can be over-loaded
while (c
i, j
> ε) do
// In case M
orig
is over-loaded
Calculate l
min
i, j
// load of the lightest task in
M
orig
for (M
dest
in all host j
in all cluster i
sorted by their credits) do
if ((M
orig
̸= M
dest
) and (c
i
, j
+ l
min
i, j
¡ ε)
then
Migrate task from M
orig
to M
dest
.
end if
end for
end while
end if
end for
taneous power consumption and energy consumption
of tasks.
5.1 Simulation Environment
Grid : 100 clusters of 100 hosts each. Each host
speed is between 1GHz and 3.06GHz
Jobs : 1000 randomly generated tasks
Duration between 10 and 40s
Requested load between 10% and 100%
Host shutdown and wakeup energy are assumed to
be zero as they are fast compared to the execution
time of tasks.
Hosts have two different power states for each
core: Switched on and switched off. While
switched on, power consumption is linear in func-
tion of load between P
min
and P
max
. Those values
are different for each host and are respectively be-
tween 75 and 150W, and 200 and 250W.
In the following we compare EACAB with other
algorithms.
Dynamic First Fit: a dynamic First Fit where
host are sorted according to their maximum power
consumption.
5.2 Experimental Results
It is designed to be a centralized coud scheduler that
emphasizes on cloud scheduler interoperation, and
complemented by a dynamic resource discovery ap-
proach on centralized network.
In this subsection, we describe the simulation
study performed to evaluate the performance of our
algorithms in terms of energy minimization as well as
the execution time and the number of migrations.
Figure 2: Energy of algorithms compared to dynamic first
fit with unsorted hosts. Lower is better.
5.3 Algorithm Description
The first observation is that for two algorithms, EA-
CAB consumes the least energy while Dynamic First
Fit algorithm consumes the most energy (see figure
2), when the number of jobs T > 350. For a small
number of tasks our algorithm leads to a significant
energy consumption. Our EACAB algorithm per-
forms even better. The second observation is that EA-
CAB algorithm is able to reduce the energy consump-
tion by 5 percent to 20 percent when job increases
from 350 to 1000.
Figures 3 and 4 show respectively the maximum
and median number of switched on host as a func-
tion of task number. When jobs increase then the
number of nodes switched on also increase, leading
to a higher power consumption. This is particularly
true if there is no migration after the initial placement
of tasks. Hence, the gain of our algorithm increases
EnergyAwareCloudsSchedulingUsingAnti-loadBalancingAlgorithm-EACAB
87
Table 1: Makespane of algorithms compared to First Fit with sorted hosts.
Number of Jobs
Algorithm 100 200 300 400 500 600 700 800 900 1000
Dynamic first Fit 265.53 374.4 580.37 631.22 615.04 870.55 813.83 1006.31 853.68 1270.73
EACAB 230 225.81 321.95 399.45 495.09 589.84 565.48 872.3 738.43 815.65
Figure 3: Maximum host switched on with EACAB.
Figure 4: Median host switched on with EACAB.
power-wise with the number of tasks because migra-
tion is activated. Computing resources are fully used
in both cases at the start of the experiment. In the
case of the consolidation, less hosts are switched-on
because we can adapt to the workload dynamism. The
median has the same behavior but the maximum num-
ber of hosts is 50%. The observed gain increases with
the number of tasks and becomes constant when hosts
are saturated.
The good results of EACAB comes from the fact
that with the increase of the number of tasks, it has
more possibilities to migrate tasks. It can then better
allocate tasks on computing resources, reducing the
number of switched-on hosts.
Figure 5: Job mean load with EACAB.
Due to the thresholds of EACAB, it would be pos-
sible to reduce further the number of switched on
hosts but it would overload remaining hosts. Those
hosts would become hot points and would have a neg-
ative impact on cooling. In order to prevent overload-
ing, EACAB adjusts load as shown in Figure 5. If
the number of tasks increases, it will reduce the mean
actual load they will obtain.
Figure 6: Variation of threshold vs energy gain with EA-
CAB
Figure 6 shows that our algorithm is better than
classical Dynamic First Fit regardless of the threshold
when the number of tasks is important.
The choice of γ is still important. There is a en-
ergy consumption difference of 10% between the best
and the worst value. The worst value is 10% more ef-
ficient than dynamic first fit, the best one is 20% more
efficient.
For small number of tasks, energy consumption
increases because of the many migrations. To choose
this value it is omportant to consider the dynamisme
of the tasks. As said previously, increasing γ, reduce
energy consumption at the cost of consolidating more
and more the tasks.
EACAB has the shortest execution time when the
number of jobs increases. The result implicates that
the scheduling algorithm such as EACAB can lever-
age interconnects with migrations to achieve high per-
formance and energy efficiency. Table 1 shows that
our algorithm produces faster scheduling regardless
of the number of jobs.
6 CONCLUSION
In this paper, we presented and evaluated our energy-
efficient migration algorithm for clouds. This algo-
rithm is based on the principle of Anti-load balancing.
SMARTGREENS2014-3rdInternationalConferenceonSmartGridsandGreenITSystems
88
It provides energy-efficiency improvement compared
to classical load unbalancing algorithms.
Our main problem was to optimize energy con-
sumption given task performance constraints. Energy
consumption is to be taken in a broad way as we try
to prevent hot spots to reduce impact on cooling.
We have compared EACAB to classical solutions
over a range of problem instances using simulation.
EACAB parameters lead to a family of heuristics that
perform well in terms of energy savings while still
leading to good task performance.
EACAB consolidate tasks on a subset of the clus-
ter hosts judiciously chosen depending on the charac-
teristics and state of resources.
This algorithm has a low computational cost. It
can then be employed in practical settings. Over-
all, the proposed EACAB algorithm can compute al-
locations effectively with an important energy gain.
Experiments showed that with our algorithm we ob-
tained a 20% gain over standard algorithms. However,
it is important to investigate further how to improve
the quality of service, but also the optimization algo-
rithm.
Also current version of EACAB is centralized. We
aim at distributing this algorithm, so that each cluster
can exchange tasks, based on their respective credits.
Future version of EACAB will also take int account
other measures to compute Credit such as network
communication patterns.
REFERENCES
Adiga, N. R., Alm
´
asi, G., Almasi, G. S., Aridor, Y., Barik,
R., Beece, D., Bellofatto, R., Bhanot, G., Bickford,
R., Blumrich, M., et al. (2002). An overview of
the bluegene/l supercomputer. In Supercomputing,
ACM/IEEE 2002 Conference, pages 60–60. IEEE.
Beloglazov, A. and Buyya, R. (2010). Adaptive threshold-
based approach for energy-efficient consolidation of
virtual machines in cloud data centers. In Proceedings
of the 8th International Workshop on Middleware for
Grids, Clouds and e-Science, page 4. ACM.
Buyya, R. and Murshed, M. (2002). Gridsim: A toolkit for
the modeling and simulation of distributed resource
management and scheduling for grid computing. vol-
ume 14, pages 1175–1220. Wiley Online Library.
Calheiros, R. N., Ranjan, R., Beloglazov, A., De Rose,
C. A., and Buyya, R. (2011). Cloudsim: a toolkit for
modeling and simulation of cloud computing environ-
ments and evaluation of resource provisioning algo-
rithms. volume 41, pages 23–50. Wiley Online Li-
brary.
Chase, J. S., Anderson, D. C., Thakar, P. N., Vahdat, A. M.,
and Doyle, R. P. (2001). Managing energy and server
resources in hosting centers. In ACM SIGOPS Op-
erating Systems Review, volume 35, pages 103–116.
ACM.
Chow, K.-P. and Kwok, Y.-K. (2002). On load balancing for
distributed multiagent computing. Parallel and Dis-
tributed Systems, IEEE Transactions on, 13(8):787–
801.
Ellis, C. S. (1999). The case for higher-level power man-
agement. In Hot Topics in Operating Systems, 1999.
Proceedings of the Seventh Workshop on, pages 162–
167. IEEE.
Hermenier, F., Lorca, X., Menaud, J.-M., Muller, G., and
Lawall, J. (2009). Entropy: a consolidation man-
ager for clusters. In Proceedings of the 2009 ACM
SIGPLAN/SIGOPS international conference on Vir-
tual execution environments, pages 41–50. ACM.
Jeon, H., Lee, W. H., and Chung, S. W. (2010). Load unbal-
ancing strategy for multicore embedded processors.
volume 59, pages 1434–1440. IEEE.
Lawson, B. and Smirni, E. (2005). Power-aware resource
allocation in high-end systems via online simulation.
In Proceedings of the 19th annual international con-
ference on Supercomputing, ICS ’05, pages 229–238,
New York, NY, USA. ACM.
Pierson, J.-M. and Casanova, H. (2011). On the utility of
dvfs for power-aware job placement in clusters. In
Euro-Par 2011 Parallel Processing, pages 255–266.
Springer.
Pinheiro, E., Bianchini, R., Carrera, E. V., and Heath,
T. (2001). Load balancing and unbalancing for
power and performance in cluster-based systems. In
Workshop on compilers and operating systems for
low power, volume 180, pages 182–195. Barcelona,
Spain.
Srikantaiah, S., Kansal, A., and Zhao, F. (2008). Energy
aware consolidation for cloud computing. In Proceed-
ings of the 2008 conference on Power aware comput-
ing and systems, volume 10. USENIX Association.
Thiam, C. and Da Costa, G. (2011). Anti-Load Balancing
to Reduce Energy Consumption (student paper). In
Ivnyi, P. and Topping, B. H. V., editors, Interna-
tional Conference on Parallel, Distributed, Grid and
Cloud Computing for Engineering, Ajaccio-Corsica-
France, 13/01/2011-15/04/2011, page (on line),
http://www.civil-comp.com/conf/progp2011.htm.
Civil-Comp Proceedings.
Warren, M., Weigle, E., and Feng, W. (2002). High-density
computing: A 240-node beowulf in one cubic meter.
In Supercomputing 2002.
EnergyAwareCloudsSchedulingUsingAnti-loadBalancingAlgorithm-EACAB
89