A Decentralized Algorithm to Revisit the Debate of Centralization and

Decentralization Approaches for Cloud Scheduling

Cheikhou Thiam

, Georges Da Costa

and Jean-Marc Pierson

Universit

e de Thi

es, Thi

es, Senegal

IRIT, Toulouse Computer Science Institute, Toulouse University, France

Keywords:

Energy, Heuristic, Virtual Machines, Cloud federation, Migration.

Abstract:

In Cloud Computing, scheduling jobs is a major and difﬁcult issue. The main objectives of cloud services

providers are the efﬁcient use of their computing resources. Existing cloud management systems are mostly

based on centralized architectures and energy management mechanisms are suffering several limitations. To

address these limitations, our contribution is to design, implement, and evaluate a novel cloud management

system which provides a holistic energy-efﬁcient VM management solution by integrating advanced VM man-

agement mechanisms such as underload mitigation, VM consolidation, and power management. In this paper,

we introduce a distributed task scheduling algorithm for Clouds that enables to schedule VMs cooperatively

and dynamically inside a federation of clouds. We evaluated our prototype through simulations, to compare

our decentralized approach with a centralized one. Our results showed that the proposed scheduler is very

reactive.

1 INTRODUCTION

Nowadays, most of the design and implementation

of energy-efﬁcient cloud management systems are

mostly based on centralized architectures. To over-

come issues bottlenecks such as overloading, under

loading and impractical unique administrative man-

agement, which are normally led by conventional

centralized or hierarchical schemes, the distributed

scheduling scheme is emerging as a promising ap-

proach because of its capability with regards to scal-

ability and ﬂexibility particularly for federation of

clouds. Energy management mechanisms are suffer-

ing several limitations. Many studies have been made

precisely to identify the parameters with the largest

impact on the total energy consumption (Khosravi

et al., 2017) (Tseng et al., 2017) and to distribute

dynamically the jobs (Brant et al., 2017) (Rouzaud-

Cornabas, 2010) (Yazir et al., 2010) (Hermenier et al.,

2010). In this paper, we introduce a distributed task

scheduling algorithm for Cloud that enables to sched-

ule and manage VMs cooperatively and dynamically

in federation clouds. Our contribution is also to de-

sign, implement, and evaluate a novel cloud man-

agement system which provides a holistic energy-

efﬁcient point of view which avoids resource overpro-

visioning. Decentralized algorithms solve the main

shortcomings of centralized algorithms such as scala-

bility, fault tolerance and bottlenecks which can sig-

niﬁcantly degrade performance, the adequacy of the

cloud computing environment and autonomy. In cen-

tralized scheduling, one cloud scheduler maintains a

complete control over the clusters. All the jobs are

submitted through the cloud scheduler. In contrast,

in decentralized scheduling, organizations maintain

(limited) control over their schedules. Jobs are sub-

mitted locally, but they can be migrated to another

cluster, if the local cluster is overloaded. The possi-

bilities of migration are, however, limited, so that mi-

grated jobs do not overload the network and the node

themselves. The aim of this article is to compare en-

ergy consumed by centralized algorithm and decen-

tralized algorithm. In this paper, we compare both

classes of scheduling algorithms, centralized and de-

centralized ones. This article is structured as follows

: Section 1 introduces the design principles. Sec-

tion 2 presents severel studies in the domain. Sec-

tion 3 presents the decentralized approach. Section

4 presents results, i.e. a comparison of a central-

ized algorithm (Energy aware clouds scheduling us-

ing anti load-balancing algorithm (EACAB)(Thiam

et al., 2014) named here Centralized) and the Decen-

tralized algorithms. Finally, Section 5 summarizes the

contribution and proposes future works.

Thiam, C., Da Costa, G. and Pierson, J.

A Decentralized Algorithm to Revisit the Debate of Centralization and Decentralization Approaches for Cloud Scheduling.

DOI: 10.5220/0006682901650172

In Proceedings of the 7th International Conference on Smart Cities and Green ICT Systems (SMARTGREENS 2018), pages 165-172

ISBN: 978-989-758-292-9

165

2 RELATED WORK

The decentralized meta-scheduling scheme allows

each cluster to own a meta-scheduler that receives

job submissions originated by local users, and to as-

sign such jobs to the local resource management sys-

tem, i.e., local scheduler. Besides the issue of ef-

ﬁciency and overhead, the Decentralized algorithm

scheme brings better scalability, compared to other

scheduling schemes. Distributed VM consolidation

algorithms enable the natural scaling of the system

when new nodes are added, which is essential for

large-scale Cloud providers (Pattanaik, 2016). An-

other beneﬁt of distributing VM consolidation algo-

rithms is the improved fault tolerance by eliminating

single points of failure: even if a server or controller

node fails, it would not render the whole system in-

operable. The dynamic scheduling of a large num-

ber of VMs as part of a large distributed infrastruc-

ture is subject to important and hard scalability prob-

lems that become even worse when VM image trans-

fers have to be managed. Consequently, most cur-

rent algorithm schedule VMs statically using a cen-

tralized control strategy. An analysis of virtual infras-

tructure managers (VIMs) reveals that most of them

still schedule VMs in a static manner. Indeed, the

advanced solutions that promote the implementation

of dynamic VM scheduling (Khosravi et al., 2017)

are subject to strong limitations regarding scalabil-

ity, reactivity, and fault-tolerance aspects. The choice

of a single master node leads to several problems.

First, during the computation and the application of

a schedule, this manager does not enforce QoS prop-

erties anymore, and thus cannot react quickly to QoS

violations. Second, because the manipulation of VMs

is costly, the time needed to apply a new schedule is

particularly important: the longer the reconﬁguration

process, the higher the risk that the schedule may be

outdated, due to the workload ﬂuctuations, when it is

eventually applied. Finally, a single master node can

lead to well-known fault-tolerance issues or a node

can be overloaded; a subgroup of VMs may be tem-

porarily isolated from the master node in case of a

network disconnection; QoS properties may not be

ensured any more if the master node crashes. Some

nodes could be overloaded which increases the en-

ergy consumed. Even if a better design of the mas-

ter can help to separate each phase in order to resolve

some of these issues (a multi-threaded model, for in-

stance, can help to track workload changes so that a

scheduling process can be stopped if needed), a cen-

tralized approach will always be subject to scalabil-

ity, reactivity, and fault-tolerance issues. In this pa-

per, we investigate whether a more decentralized al-

gorithm approach can tackle the aforementioned lim-

itations. Indeed, scheduling takes less time if the work

is distributed among several nodes, and the failure

of a node does not impede scheduling strongly any

more. Several proposals have been made precisely

to distribute dynamic VM management (Yin et al.,

2016) (Yang et al., 2016) (Feller et al., 2010) (Yazir

et al., 2010). However, the resulting prototypes are

still partially centralized. Firstly, at least one node

has access to a global view of the system. Secondly,

several virtual infrastructure managers (VIMs) con-

sider all nodes for scheduling, which limits scalabil-

ity. Thirdly, several VIMs still rely on service nodes,

potential single points of failure. (Quesnel and L

ebre,

2011) introduces distributed VM scheduler (DVMS),

a VIM that schedules and manages VMs coopera-

tively and dynamically in distributed systems. Author

designed it to be nonpredictive and event-driven, to

work with partial views of the system, without any po-

tential single points of failure. Our VIM thus has the

same characteristics and is more reactive, more scal-

able, and more tolerant to nodes crashes or network

disconnections. Kang and Choo (Kang and Choo,

2016) introduced an Inter Cloud Manager (ICM) job

dispatching decentralized algorithm which operates

well in large scale environments. Authors in (Mehr-

sai et al., 2017) propose a hybrid algorithm which is a

coalition between both techniques (hybrid centralized

and decentralized). The proposed approach can effec-

tively adhere to strict QoS requirements, as well as

handle multi-core CPU architectures, heterogeneous

infrastructure and heterogeneous VMs.

3 A DECENTRALIZED TASK

SCHEDULING ALGORITHM

FOR CLOUD

The algorithm uses a multi-phase decentralized al-

gorithm scheduling solution. It is comprised of a

migration phase, a job submission phase responsi-

ble for job dissemination, as well as the dynamic

scheduling phase responsible for iterative scheduling

improvement. Furthermore, the decentralized algo-

rithm is a collection of implementation independent

interfaces and heuristics, which are used to facilitate

job scheduling across distributed nodes.

3.1 Algorithm Statements

There are in total N sites and in each site a set H

nodes distributed in a cloud data center system

with the same start time U. H

i, j

is the node j in site

i. Each time when one node (initiator) attempts to

SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems

166

(re)assign a task to another node (or the same node)

for execution, the initiator is called the requester node,

and the node receiving such a request is called the re-

sponder node. Each task V M

i, j,k

i ∈ [1,2,...,N] and

j ∈ [1,2,...,nH

] is sent from node H

i, j

. The decen-

tralized algorithm is comprised of two phases, namely

the job submission phase and the dynamic scheduling

phase, which work together to ensure both a quick job

distribution and an optimized rescheduling effect.

• Job submission phase. This phase is the ﬁrst phase

of the algorithm. Each time a node j, receives

a V M

i, j,k

submitted by its local user, node j be-

haves as a requester node H

req

i, j

and generates a

request message V M

req

i, j,k

for V M

i, j,k

. VM char-

acteristic information including estimated execu-

tion time D

i, j,k

and requested amount of Proces-

sor Elements (PEs) V M

i, j,k

will be appended to

the generated request message. Afterwards, re-

quest message V M

req

i, j,k

is replicated and dissem-

inated to each of the discovered remote nodes

asking for the job delegation possibilities. All

nodes receiving the job delegation request mes-

sage V M

req

i, j,k

, including the requester node j itself,

are considered as responder nodes. Each respon-

der node H

rep

, j

needs to send an accept message

V M

ack

i, j,k

whether the node j’ is able and willing

to execute the received V M

i, j,k

. A V M

req

i, j,k

takes

various factors, such as the current state of the

node which allows it to execute the VM using

the most efﬁcient amount of energy relative to

other nodes. If yes, each candidate, considered

as responder node, computes an estimated com-

pletion time according to its current scheduling,

mainly the energy consumed and resource status

and delivers the information by means of an ac-

cept message V M

ack

i, j,k

. In addition, the estimated

response time and the estimated energy consumed

if V M

i, j,k

is executed by node j’ in site i’ are also

appended to the generated accept message, which

can be utilized by the requester node for respon-

der node evaluation and selection. Each time a

request message V M

req

i, j,k

is generated by the re-

quester node and disseminated to contactable re-

mote nodes, the requester node waits and collects

all received VM

ack

i, j,k

, generates an assign message

V M

ass

i, j,k

and send it to a proper remote node H

ass

, j

(assignee node) to which the VM

i, j,k

is delegated.

An arbitrary node j in site i, due to the effect of

the job submission phase, has received a set of

jobs from either local users’ submissions or re-

mote nodes’ delegations. A rescheduling process

helps this approach to adapt to the changes of both

underlying resources and arriving jobs.

• Dynamic scheduling phase. This phase solves

some problems related to the ever changing

data center infrastructure during VM submission

phase. It allows for example a redistribution al-

gorithm for a VM that is in a long tail and thus a

node can not be executed instantly. Suppose the

arbitrary node H

i, j

ﬁnds that many customers ap-

plications (VMs) wait to be served, such selected

customers will be sent to others nodes (reschedul-

ing) to improve their own VM makespans and the

overall resource utilization, therefore reducing the

energy consumed of the overall cloud data cen-

ter. The number of to-reschedule VMs is com-

puted by considering the status of node H

i, j

it-

self, such as the length of local waiting customers

and current node overhead. Afterwards, the al-

gorithm ﬁnds a set of contactable remote nodes

for VM rescheduling and re-allocation. H

i, j

sends

an inform message V M

in f o

i, j,k

for each to-schedule

V M

i, j,k

, and disseminates those inform messages

to all remote nodes discovered for negotiating VM

rescheduling possibilities. Each generated mes-

sage V M

in f o

i, j,k

, contains ﬁrstly the same informa-

tion as the aforementioned request message, in-

cluding estimated execution time D

i, j,k

, requested

amount of PEs V M

i, j,k

, and job characteristic pro-

ﬁle. Furthermore, each inform message also in-

cludes the already made schedule for V M

i, j,k

, on

the current node H

i, j

, i.e., the estimated VM ﬁnish

time and the estimated energy consumed, which

will be used for offer comparison by the contacted

remote nodes later. It is noteworthy that the al-

gorithm can be triggered by customized events as

well depending on each node local setting. The

selection of the node that will receive the task is

the same as during the submission phase except

that the initiator is no longer a candidate node.

3.2 Scenario

Assuming that a cloud is comprised of interconnected

nodes H

i, j

, i ∈ [1,2,...,N] and j ∈ [1,2,...,nH

V M

i, j,k

is submitted to node H

i, j

. The Decentralized

algorithm then leads to the following phases and steps

for job assignment and dynamic scheduling. When a

new VM arrives there is a VM submission phase.

• VM Submission Phase.

– Step 1 : V M

i, j,k

is submitted to node H

req

i, j

to-

gether with its characteristic data.

– Step 2 : The initiator node H

req

i, j

generates a

request message according to the retrieved ex-

ecution requirement from V M

i, j,k

, and broad-

casts such a message to the cloud by the em-

A Decentralized Algorithm to Revisit the Debate of Centralization and Decentralization Approaches for Cloud Scheduling

167

ployed lightweight decentralized algorithm in-

formation system.

– Step 3 : The initiator node H

req

i, j

then waits for

some time to receive returned accept messages

from other nodes of the data center.

– Step 4 : Nodes receiving the broadcast request

message check if the required VM proﬁle can

be matched by the local resources.

– Step 5 : If yes, each candidate node computes

an estimated VM response time according to its

current scheduling, their loads (load increases

while staying between the under-loaded thresh-

old and the over-loaded threshold) and resource

status, and sends the information back to the

initiator node H

req

i, j

by means of an accept mes-

sage.

– Step 6 : The initiator node H

req

i, j

evaluates re-

ceived accept messages and selects the can-

didate node which fulﬁlled requirements and

minimize the energy consumption. The se-

lected node, referred to the assignee node, is

assigned the VM by means of an assign mes-

sage. A node H

ass

, j

is selected as the candidate

node for V M

i, j,k

delegation.

Regularly the system tries to schedule the waiting

VM.

• Dynamic Scheduling Phase.

– Step 7 : The assignee node H

ass

, j

takes the as-

signed V M

i, j,k

and puts it into its local list of

VMs.

– Step 8 : The assignee node H

ass

, j

periodically

picks jobs from its local list of VMs, which

have a long enough waiting time and have not

been selected recently. Afterward, for each se-

lected VM, the assignee node H

ass

, j

generates an

inform message, which contains both VM es-

timated completion time and energy estimated

for VM execution. In our case, V M

i, j,k

is se-

lected with an assumed long waiting time.

– Step 9 : The inform message of V M

i, j,k

is sent

over the network using a low-overhead walking

protocol (the walk starts at an initial node and at

each step selects randomly a minimum number

of neighbors).

– Step 10 : Nodes receiving the aforementioned

inform message check if the local resource and

scheduling status could match the requirement

of V M

i, j,k

; furthermore, they also need to check

whether the estimated VM response time is

short enough when compared to V M

i, j,k

current

response time upon the assignee node H

ass

, j

– Step 11 : If the evaluation result from above

step is positive, an accept message will be gen-

erated and delivered to the assignee node H

ass

, j

– Step 12 : The assignee node H

ass

, j

evaluates

the received accept messages according to the

promised VM response time. As a result, node

re−ass

, j

is selected, and V M

i, j,k

is re-assigned

by means of an assign message.

– Step 13 : To enable tracking of VMs for the

purpose of node crash tolerance, each VM re-

assignment is logged and notiﬁed to the initia-

tor node H

req

i, j

– Step 14 : To enable node weighting for future

scheduling, VM completion status is sent back

to the original assignee node H

ass

, j

and initiator

node H

req

i, j

Regularly the scheduling restarts to optimize the

metrics.

• Migration Phase. Two steps :

– Step 1 : For a node H

i, j

, if c

i, j

> ε the node is

considered as being overloaded (with c

i, j

rep-

resenting the load of the node j in cluster i);

Node H

i, j

requires migration of one or more

of its VMs. The algorithm must determine

what part of the current load should be sent to

other nodes. The algorithm seeks, for all nodes

in all sites, one or more nodes, whose charge

γ < c

, j

< ε, which can receive the VM to mi-

grate. If such node is found, the execution of

the VM is stopped at the source node H

i, j

and

continues at the destinations node. If the decen-

tralized algorithm is to add one or more nodes,

the algorithm must determine what part of the

current load should be sent to the added nodes.

Obviously, the load to be migrated should come

from nodes undergoing excessive demand for

resources.

– Step 2 : if c

i, j

< γ the node H

i, j

is considered

as being underloaded; Node H

i, j

requires mi-

gration of its VMs. The algorithm seeks, for

all nodes in all sites, one or more nodes whose

charge γ < c

, j

< ε which can receive the load

to migrate. If such nodes is found, the execu-

tion of all VMs is stopped at the source node

i, j

and continues at the destinations node.

3.3 Algorithms

Our Energy Efﬁcient Decentralized Scheduling for

Cloud adopts the promised VM response time and the

node load as main criterions to evaluate the nodes’ ca-

pabilities. Participating nodes need to calculate their

SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems

168

actual load and estimated response time for a con-

cerned VM and bid for the VM delegation using the

calculated and promised VM response time. Each re-

sponder node computes an estimated completion time

according to its current scheduling and resource sta-

tus, calculates the necessary energy, and delivers the

information by means of an accept message. In addi-

tion, the node load is also added to the generated mes-

sage, which can be utilized by the requester node for

responder node evaluation and selection. The selected

node is the best candidate node based on several pa-

rameters, such as the promised time to complete, en-

ergy consumed, the node load between under-load

threshold and over-load threshold, node weight due to

historical interaction records, etc. Furthermore, dur-

ing the execution of the VMs, all the time the system

checks if there are underloaded or overloaded nodes.

A self-healing is designed similar to the intelligent

feedback loop, and supports not only scheduling but

also re-scheduling activities. It enables Decentralized

algorithm with self-healing capabilities to allow VMs

that wait a long time in a node to be re-scheduled.

The dynamic scheduling phase, but also the possibil-

ity to switch off or switch on nodes, and especially

the cooperation between nodes help to maintain sys-

tem efﬁciency. These properties of Decentralized al-

gorithm constitute its self-healing system. Decen-

tralized algorithm provides task scheduling strategy,

which dynamically migrate VMs among computing

nodes, transferring VMs from underloaded nodes to

loaded but not overloaded nodes. It balances load of

computing nodes as far as possible in order to reduce

program running time. The Decentralized algorithm

behaves globally as follows:

• If total VM load on the node j of site i > ε, node

is over-loaded. Let ε be the over-load threshold.

• If total VM load on the node j of site i < γ, node is

under-loaded. Let γ be the under-load threshold.

The algorithm is described in Algorithm 4. For

the sake of simplicity, corner cases such as all nodes

over-loaded are not included. Selection policies take

into account migration cost. First the algorithm com-

putes the current load c

i, j

(line 6) of the host. Then it

calls the CheckLoad procedure (see Algorithm 3) that

will either migrate all VMs (if node is underloaded,

in order to switch H

i, j

off) or some VMs (if node

is overloaded, so that it becomes again in normal

mode). Then the algorithm checks if some VM are

still waiting for execution. If so, it selects the one

waiting for the longest period (see Algorithm 1) and

migrate it to the ﬁrst host where it can enter, hosts

being sorted by their maximum power consumption

max

Algorithm 1 : Find VM with longest remaining waiting

time: FindVmWait.

Input: V M

i, j

// set of VM of node j of site i

Output: V M

i, j,k

// V M

i, j,k

in node j of site i

1: function FINDVMWAIT(x)

2: for V M

i, j,k

in V M

i, j

3: D = 0

4: if (D

i, j,k

> D) then // D

i, j,k

is the waiting

time of VM k in node j of site i

5: D= D

i, j,k

6: k’ = k

7: end if

8: end for

9: Return V M

i, j,k

10: end function

Algorithm 2: Find VM with longest remaining execution

time: FindVmExec.

Input: V M

i, j

//set of VM of node j of site i

Output: V M

i, j,k

// V M

i, j,k

in node j of site i

1: function FINDVMEXEC(V M

i, j

)

2: for V M

i, j,k

in V M

i, j

3: D = 0

4: if (D

Exec

i, j,k

> D) then // D

Exec

i, j,k

is the execu-

tion time of VM k in node j of site i

5: D= D

Exec

i, j,k

6: k’ = k

7: end if

8: end for

9: Return V M

i, j,k

10: end function

The procedure CheckLoad() (described in Algo-

rithm 3) is composed of two parts. In case of under-

load (line 4 to 10), it looks for potential destinations

for all present VMs. When many potential destina-

tions exist, the one with the minimum migration cost

is selected. In case of overload, the algorithm repeat-

edly looks for the VM with maximum remaining exe-

cution time (see Algorithm 2), and migrates this VM

to the ﬁrst node that could host it (sorted by P

max

until the load is under the overload threshold ε. In

both cases, if no destination is found, the VM remains

where it is (and the node remains either underloaded

or overloaded).

3.4 Analysis of the Algorithm

The decentralized algorithm is inspired by the moti-

vation of enabling cloud scheduling for the scope of

the overall cloud, instead of each single node. In con-

trast to conventional cloud scheduling solutions, this

A Decentralized Algorithm to Revisit the Debate of Centralization and Decentralization Approaches for Cloud Scheduling

169

Algorithm 3: Migration of VM based on sorted nodes.

1: procedure CHECKLOAD(c

i, j

)

Input: H

i, j

// the host running this procedure

2: set Potential =

3: set H = {H

, j

∈ [1, 2, ..., N], j

∈

[1,2,...,nH

],i

6= i, j

6= j}, sorted by P

max

4: if c

i, j

< γ then

5: for (H

, j

in H) do

6: if c

i, j

+ c

, j

< ε then

7: Add H

, j

to Potential

8: end if

9: end for

10: Migrate all VMs from H

i, j

to the node in

Potential with minimum migration cost

11: else

12: while (c

i, j

> ε) do

13: V M

i, j,k

= FindV MExec(V M

i, j

)

14: l

i, j,k

= V M

i, j,k

load

15: for (H

, j

in H) do

16: if c

, j

+ l

i, j,k

< ε then

17: Migrate V M

i, j,k

from H

i, j

, j

18: end if

19: end for

20: end while

21: end if

22: end procedure

Algorithm 4: Decentralized scheduling.

Input: H

i, j

// the host running this procedure

1: // P

min

, minimum power

2: // P

max

, maximum power

3: procedure DECENTRALIZEDSCHEDULING

4: for Regularly do

5: set H = {H

, j

∈ [1,2,...,N], j

∈

[1,2,...,nH

],i

6= i, j

6= j}, sorted by P

max

6: Compute c

i, j

// it is the load of H

i, j

7: CheckLoad(c

i, j

)

8: if (notempty V M

wait

i, j

) then

9: V M

i, j,k

=FindVMWait(V M

wait

i, j

)

10: l

i, j,k

= V M

i, j,k

load

11: for H

, j

in H do

12: if (c

, j

+ l

i, j,k

) < ε then

13: Migrate V M

i, j,k

from H

i, j

, j

14: end if

15: end for

16: end if

17: end for

18: end procedure

algorithm is designed to deliver relevant scheduling

events, such as VM submission and scheduled in-

formation, to as many neighboring remote nodes as

possible. Moreover, the algorithm enables the pos-

sibility of dynamic rescheduling in order to adapt

to cloud data center characteristics such as instan-

taneity and volatility. Our decentralized algorithm

also uses consolidation in order to minimize energy

consumed. The algorithm is comprised of three

phases, namely the VM submission phase, dynamic

rescheduling phase and migration phase. The VM re-

sponse time and the energy are used as critical criteria

to evaluate the energy consumed, the performance of

scheduling and rescheduling. The decentralized algo-

rithm is able to work together with different kinds of

local scheduling algorithms.

4 RESULTS

We run the simulation using a Grid5000 workload and

measure the amount of energy consumed by the place-

ment and the average execution times for both algo-

rithms (i.e. centralized and Decentralized algorithm).

To derive the actual energy savings, the amount of en-

ergy spent for VM placement was estimated by the

formula deﬁned below.

Nodes have two different power states : Switched on

and switched off. While switched on, power con-

sumption is linear in function of load between P

min

i, j

and P

max

i, j

We use the classical linear model of power con-

sumption in function of load :

∀i, j P

i, j

= P

min

i, j

max

i, j

-P

min

i, j

)

if node j is switched on, P

i, j

= 0 otherwise.

Therefore the total power consumption of the system

is:

P =

∑

i=1

∑

j=1

i, j

To obtain energy consumed during a time slice, in-

stantaneous power is integrated over time

P(t)dt.

Total energy is then obtained by summing all the en-

ergy of those time slices.

The ﬁnal simulation results show the gain in en-

ergy and makespan does not depend on the numbers

of jobs but mostly of the distribution of jobs between

nodes. The ﬁnal simulation results are depicted in ta-

ble 1. The gain in energy and makespan don’t depend

on the number of jobs but mostly of the distribution

of jobs between nodes.

Figure 1 demonstrates the number of migrations

incurred by 4500 jobs. An obvious observation is

that migrations are beneﬁcial to save energy. The

second phase is the dynamic scheduling phase. If at

any time the tasks are ﬁnished and there are several

SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems

170

Table 1: Simulation results (StDev= standard deviation).

Job Approach Makespan Makespan

gain

Makespan

StDev

Energy Energy

gain

energy

StDev

500

Cent 48770 6491 16358.5 10363419

Dec 48984 -0.44 5574 16252.9 0.65 10475850

1000

Cent 102089 40225 34473.4 56793603

Dec 100859 1.20 40676 31628.8 8.25 51366829

1500

Cent 117143 24348 45480.4 104549039

Dec 130734 -11.6 61778 47263.9 -3.9 118552002

2000

Cent 153386 51215 58819.5 154963019

Dec 143047 6.74 39176 63349.97 7.7 180703187

2500

Cent 143977 37093 56716.3 129407887

Dec 143541 0.3 36334 56108.85 1.07 116130251

3000

Cent 250457 237465 61635.4 167952610

Dec 211032 15.72 191168 56993.7 7.5 133578369

3500

Cent 235614 215901 63370.7 164512761

Dec 202696 13.97 162140 55093.6 13.06 130973378

4000

Cent 221526 192048 54071.2 116554623

Dec 233452 5.38 223754 60270.2 11.5 157261329

4500

Cent 241048 222470 57983.4 131320566

Dec 207189 14.05 173065 59478.3 2.57 175986327

machines under loaded, they can be consolidated.

This is not the case for algorithms which never calls

into question the allocation once the jobs are running.

The impact of migration, however, may not be large

enough to dominate the total energy gain when the

number of jobs is less than 300. We performed series

of tests, comparing Centralized and Decentralized

algorithm.

The Centralized algorithm (EACAB) (Thiam

et al., 2014) 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.

Generally, the Decentralized algorithm schedules

slightly worse in terms of makespan than the Central-

ized algorithm.

Figure 1: Centralized (Cent) vs Decentralized (Dec) algo-

rithms: Energy.

Typical results of experiments are presented in

Figures 1, 2, 3 and 4. In Figure 1 it is easy to notice

that energy consumed by the distributed algorithm is

comparable to centralized strategy for low number

of jobs. These poor results are caused by low num-

ber of migrations since majority of jobs can be exe-

Figure 2: Centralized (Cent) vs Decentralized (Dec) algo-

rithms: Migration.

Figure 3: Centralized (Cent) vs Decentralized (Dec) algo-

rithms: Maximum nodes switched on.

Figure 4: Centralized (Cent) vs Decentralized (Dec) algo-

rithms: Makespan.

cuted without exceeding their due dates. This situa-

tion changes for higher loads when number of migra-

tions is increased and the distributed algorithm out-

performs the centralized in some case. We achieved

similar results for centralized algorithm which use mi-

gration and anti load-balancing techniques. When

a cluster load is below the under-loaded threshold,

centralized and Decentralized algorithm are able to

migrate jobs to more-loaded clusters and switch off

under-loaded cluster. In this case, performance mea-

sures and energy depend strongly on the collabora-

tion of less-loaded clusters. When their cooperation is

too low the system as a whole starts to be inefﬁcient,

although the performance of the less-loaded clusters

is not affected. Consequently, we consider that there

must be some minimal cooperation that results from

a cloud agreement. As in real systems the job stream

changes, this minimal cooperation can be also inter-

A Decentralized Algorithm to Revisit the Debate of Centralization and Decentralization Approaches for Cloud Scheduling

171

preted as an ”insurance” to imbalance the load. From

the experiments above, we can get the obvious con-

clusion that both the Centralized algorithm and De-

centralized algorithm can reduce energy consumed of

data centers. Figure 4 above shows the execution time

for all tasks and both schedulers. We can see that the

two algorithms have the same behavior.

5 CONCLUSIONS

We have compared the Decentralized algorithm to

Centralized algorithm EACAB over a range of real-

istic problem instances using simulation. The pro-

posed algorithm can compute allocations effectively

with an important energy gain. The simulation re-

sults showed that our algorithm is capable of obtain-

ing energy-efﬁcient schedules using less optimiza-

tion time. Our experimental results have shown that:

(1) the Decentralized algorithm is capable of obtain-

ing energy-efﬁcient schedules using less optimization

time; (2) Application performance is not impacted by

performing migration using under load and over load

thresholds; (3) The system scales well with increasing

number of resources thus making it suitable for man-

aging large-scale virtualized data centers; (4) The anti

load-balancing technique used by the two approaches

achieve substantial energy savings.

Finally we can get the obvious conclusion that both

Centralized and Decentralized algorithm algorithm

can reduce energy consumed of data centers. Com-

pared to centralized algorithms, decentralized algo-

rithms have a simplicity that makes them promising

in practice though for a veriﬁcation more experiments

are required.

REFERENCES

Brant, C., Choudhary, S., Garrison, J., and McKay, M.

(2017). Distributed virtual machine image manage-

ment for cloud computing. US Patent 9,747,124.

Feller, E., Rilling, L., Morin, C., Lottiaux, R., and Lep-

rince, D. (2010). Snooze: a scalable, fault-tolerant and

distributed consolidation manager for large-scale clus-

ters. In Proceedings of the 2010 IEEE/ACM Int’l Con-

ference on Green Computing and Communications &

Int’l Conference on Cyber, Physical and Social Com-

puting, pages 125–132. IEEE Computer Society.

Hermenier, F., L

ebre, A., and Menaud, J.-M. (2010).

Cluster-wide context switch of virtualized jobs. In

Proceedings of the 19th ACM International Sympo-

sium on High Performance Distributed Computing,

pages 658–666. ACM.

Kang, B. and Choo, H. (2016). A cluster-based de-

centralized job dispatching for the large-scale cloud.

EURASIP Journal on Wireless Communications and

Networking, 2016(1):25.

Khosravi, A., Andrew, L. L., and Buyya, R. (2017). Dy-

namic vm placement method for minimizing energy

and carbon cost in geographically distributed cloud

data centers. IEEE Transactions on Sustainable Com-

puting, 2(2):183–196.

Mehrsai, A., Figueira, G., Santos, N., Amorim, P., and

Almada-Lobo, B. (2017). Decentralized vs. cen-

tralized sequencing in a complex job-shop schedul-

ing. In IFIP International Conference on Advances

in Production Management Systems, pages 467–474.

Springer.

Pattanaik, S. (2016). Efﬁcient Energy Management in Cloud

Data center using VM Consolidation. PhD thesis.

Quesnel, F. and L

ebre, A. (2011). Cooperative dynamic

scheduling of virtual machines in distributed sys-

tems. In European Conference on Parallel Processing,

pages 457–466. Springer.

Rouzaud-Cornabas, J. (2010). A distributed and collabo-

rative dynamic load balancer for virtual machine. In

Euro-Par Workshops, pages 641–648. Springer.

Thiam, C., Da Costa, G., and Pierson, J.-M. (2014). Energy

aware clouds scheduling using anti-load balancing al-

gorithm: Eacab. In 3rd International Conference on

Smart Grids and Green IT Systems (SMARTGREENS

2014), pages pp–82.

Tseng, H.-W., Yang, T.-T., Yang, K.-C., and Chen, P.-S.

(2017). An energy efﬁcient vm management scheme

with power-law characteristic in video streaming data

centers. IEEE Transactions on Parallel and Dis-

tributed Systems.

Yang, D., Deng, C., and Zhao, Z. (2016). Dynamic schedul-

ing method of virtual resources based on the pre-

diction model. In International Conference on Col-

laborative Computing: Networking, Applications and

Worksharing, pages 384–396. Springer.

Yazir, Y. O., Matthews, C., Farahbod, R., Neville, S., Gui-

touni, A., Ganti, S., and Coady, Y. (2010). Dynamic

resource allocation in computing clouds using dis-

tributed multiple criteria decision analysis. In Cloud

Computing (CLOUD), 2010 IEEE 3rd International

Conference on, pages 91–98. Ieee.

Yin, L., He, W., and Luo, J. (2016). Distributed virtual ma-

chine placement based on dependability in data cen-

ters. In Natural Computation, Fuzzy Systems and

Knowledge Discovery (ICNC-FSKD), 2016 12th In-

ternational Conference on, pages 2152–2158. IEEE.

SMARTGREENS 2018 - 7th International Conference on Smart Cities and Green ICT Systems

172