0% found this document useful (0 votes)

40 views15 pages

Dynamically Partitioning Workflow Over Federated Clouds For Optimising The Monetary Cost and Handling Run-Time Failures

This document proposes a framework called DoFCF that can dynamically partition scientific workflows across multiple public and private cloud data centers. This is done to minimize financial costs while meeting security requirements and handling runtime failures. The framework partitions workflows by deploying sensitive services and data to more secure private clouds, and less sensitive services and data to cheaper public clouds. It was tested using a cloud simulation tool and a real cloud platform, and was able to reduce costs while adapting to failures and meeting security needs.

Uploaded by

arunkumargoge

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

0% found this document useful (0 votes)

40 views15 pages

Dynamically Partitioning Workflow Over Federated Clouds For Optimising The Monetary Cost and Handling Run-Time Failures

Uploaded by

arunkumargoge

We take content rights seriously. If you suspect this is your content, claim it here.

Available Formats

Download as PDF, TXT or read online on Scribd

You are on page 1/ 15

This article has been accepted for publication in a future issue of this journal, but has not been

fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2016.2603477, IEEE
Transactions on Cloud Computing
JOURNAL OF LATEX CLASS FILES, VOL. ??, NO. ??, ??? 1

Dynamically Partitioning Workflow over

Federated Clouds For Optimising the Monetary
Cost and Handling Run-Time Failures
Zhenyu Wen∗ ,§ , Rawaa Qasha∗‡ , Zequn Li† , Rajiv Ranjan∗ , Paul Watson∗ , Alexander Romanovsky∗

Abstract—Several real-world problems in domain of healthcare, large scale scientific simulations, and manufacturing are organised as
workflow applications. Efficiently managing workflow applications on the Cloud computing data-centres is challenging due to the
following problems: (i) they need to perform computation over sensitive data (e.g. Healthcare workflows) hence leading to additional
security and legal risks especially considering public cloud environments and (ii) the dynamism of the cloud environment can lead to
several run-time problems such as data loss and abnormal termination of workflow task due to failures of computing, storage, and
network services. To tackle above challenges, this paper proposes a novel workflow management framework call DoFCF (Deploy on
Federated Cloud Framework) that can dynamically partition scientific workflows across federated cloud (public/private) data-centres for
minimising the financial cost, adhering to security requirements, while gracefully handling run-time failures. The framework is validated
in cloud simulation tool (CloudSim) as well as in a realistic workflow-based cloud platform (e-Science Central). The results showed that
our approach is practical and is successful in meeting users security requirements and reduces overall cost, and dynamically adapts to
the run-time failures.

Index Terms—Cloud Federation, Scientific Workflow Optimisation, Deployment, Security, Monetary Cost, Scheduling.

1 I NTRODUCTION

S CIENTIFIC workflows have become an increasingly pop-

ular paradigm for enabling and accelerating scientific
data analysis. They consist of a series of computational tasks
tations mainly address a single cloud. In addition, existing
systems are unable to coordinate across different cloud-
based data centres in order to optimally allocate application
that are logically connected by data and controlling flow de- services to meet users’ functional (i.e. hardware) or non-
pendencies. They have been successfully run on traditional functional (i.e. security) requirements [5]. Further, cloud
HPC (High Performance Computing) systems and clusters. providers are subject to failures. For example, an outage
However, in recent years many researchers have migrated at GoDaddy took down millions of web sites [6] and a 12
workflow systems onto the cloud in order to exploit the hour Amazon EC2 outage [7] raised serious questions about
economic and technical benefits of this technology [1]. reliability on a single cloud provider.
The Cloud computing provides a computing paradigm Considering these issues, cloud federation has the po-
that focuses on the on-demand provisioning of comput- tential to facilitate just-in-time, opportunistic, and scalable
ing resources, including hardware, software, storage and provisioning of application services, consistently fulfilling
network. Furthermore, cloud providers have distributed user requirements under variable workload, resource and
several data centres at different geographical locations over network conditions [8]. Using a federated cloud, users
the internet in order to deliver quality services for their are able to deploy their applications over a set of clouds
customers around the world [2] [3]. The currently available from different cloud providers across different geographical
cloud platforms distinguish themselves on service type, the locations, bringing various advantages such as leveraging
cost, the Quality of Service (QoS) as well as performance [4]. unique cloud specific services, higher availability and re-
This fact enables cloud customers to freely select their target dundancy, disaster recovery and geo-presences.
architecture from a broad range of cloud platforms. Motivated by the above considerations, we propose to
Although users are interested in the execution of their deploy scientific workflow over utility-oriented cloud fed-
workflow applications in the cloud, the current implemen- erations, in the meantime ensuring that the deployment can
meet users’ security requirements and minimise monetary
∗ The authors are with the School of Computing Sci- cost, while handling changes in cloud availability (including
ence, Newcastle University, United Kingdom. E- the existing cloud virtual machines (VMs thereafter) outage
mail:{z.wen,r.qasha,Raj.Ranjan,paul.watson,alexander.romanovsky}
@newcastle.ac.uk,
and the new VMs becoming available).
† The author is with the school of mathematics and infor- We assume that there is a set of workflow execution
mation science, Northumbria university, United Kingdom. E- environments running on different data centres which are
mail:{Zequn.li}@northumbria.ac.uk, owned by different cloud providers. Further, the different
‡ The author is with the College of Computer Sciences and Mathematics,
Mosul university, Iraq. computing resources have different security levels, for ex-
§ The author is with the School of Informatics, University of Edinburgh, ample, Windows Azure provides private cloud and public
United Kingdom. E-mail:{zwen}@inf.ed.ac.uk cloud that come with different security levels. Therefore,
Manuscript received ???; revised ??? the security of a scientific workflow can be improved by

deploying sensitive services or data to more secure clouds. • Investigation of the existing state-of-the-art optimisa-
Likewise, the cost can be reduced through distributing the tion algorithms. Further, we extend two classic algo-
less sensitive services or data to cheaper clouds with lower rithms and adapt to DoFCF to achieve rapid explo-
security levels. In addition, we allow for the cloud federa- ration for a possible deployment solution. In order to
tion to be very dynamic, as the availability of clouds may handle the availability change of cloud resources, a
sometimes change. novel dynamic rescheduling algorithm is developed
According to the above assumptions, the deployment of to resume workflow execution when failures occur or
scientific workflow on a federated clouds poses a number of reduce the monetary cost by redeploying the running
challenges: workflow to cheaper clouds.
• Evaluating the implemented framework on
• The considerable amount of computation required
CloudSim [11] which is a Cloud simulator and
for exploring the optimal deployment. We assume
e-Science Central [12] (e-SC), a real scientific
a workflow with S tasks can be deployed over a
workflow based cloud platform.
federated cloud that includes C clouds. Therefore
the total number of deployments is C S which is The rest of this paper is organised as follows. In Section 2
exponential to the number of workflow tasks. the basic models of the framework are discussed. Next, a
• Tasks in the workflow system and their correspond- specific security model is adapted to DoFCF, demonstrat-
ing security levels are influenced by different aspects, ing how to deploy a workflow over a federated cloud to
such as user preference, the task requirements and meet security requirements while minimising the cost. Then
inputs/outputs data. the state-of-the-art optimisation algorithms are explored,
• The cost of the deployment is dependent on several extended and adapted to our DoFCF to optimise workflow
factors, including data storage cost, data communi- partitioning. In Section 5, we evaluate the framework by
cation cost and computation cost. using CloudSim, and also develop a tool to schedule the
• The trade-off between secure deployment and mon- workflows over a set of e-SC instances. Before drawing
etary cost is also a challenge. conclusions in Section 7, we discuss the related work.
• Dynamic handling of cloud environment changes.
This requires to rapidly generate new deployment
solutions when the cloud environments change.
2 BASIC M ODELLING C ONSTRUCTS
In this section, we present a system model of deploying
{Few of works have been done to address deploying workflow workflow applications over federated clouds. In the fol-
over federated cloud. [9] introduces a static algorithm to deploy lowing, the general scientific workflow model and security
workflows over federated clouds that meets security requirements. model are introduced. In addition, a general cost model will
However, the dynamic of cloud environments is not considered. In be used to calculate the monetary cost of deployment. More-
our previous work [10], we have considered cloud failure during over, we present an optimisation model that can guarantee
workflow running, while meeting the security requirements and the deployment solution meets the security constraints as
minimising the cost. However, the proposed method cannot handle well as minimising the cost. Finally, a dynamic cost model
the large scale workflow and fails to generate a better solution is used to help rescheduling the running workflow when
when a new cloud joins the cloud federation.} the cloud availability change in a federated clouds. Table 1
In this paper, we propose DoFCF (Deploy on Federated shows the notations for the rest of the paper.
Cloud Framework), a framework for deploying workflows
over federated clouds that meets security requirements, and
optimising the monetary cost, while dynamically handling 2.1 System model
the availability change of federated clouds. A Cloud Service Broker performs cloud exchange and ne-
Our framework provides a set of solutions for work- gotiates with each available cloud to allocate resources that
flow scheduling, including where to deploy tasks, when to meet user’ specific requirements. In this paper, we propose
start each service and how to handle the cloud availability a Cloud Service Broker which can partition workflow appli-
change. The deployment of the workflow over federated cations over federated clouds. Fig 1 shows the architecture
cloud is based on adhering to a set of specific security of the Cloud Service Broker for workflow application along
requirements and minimisation procedures. Additionally, with other components as illustrated below:
DoFCF offers a dynamic solution to dynamically reschedule The Client can be a platform for workflow management
the running workflow to new clouds, in order to complete such as e-Science Central or Pegasus [13] which allows users
the execution of an unfinished workflow or save on costs. to describe and submit their workflow application through
platform components. The Workflow Engine delivers the
1.1 Paper Contributions workflow tasks (or services in this paper) to the underlying
Cloud Service Broker, including execution requirements,
Considering the above challenges and problems, this paper
task description, and the desired security requirements.
makes the following core contributions:
The Cloud Service Broker enables the functions of resource
• A framework to model the security constraint of allocation, workflow scheduling and software deployment.
workflow deployment and the situation of cloud Our framework includes a Planner component that per-
availability change during the workflow execution forms a matching process to select the target clouds for
time. The framework also quantifies the cost of exe- deployment, based on the information passed from Global
cuting workflow over federated clouds. Cloud Market. Further, the workflow tasks are assigned by

2168-7161 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2016.2603477, IEEE
Transactions on Cloud Computing
JOURNAL OF LATEX CLASS FILES, VOL. ??, NO. ??, ??? 3
Symbol Meaning
Workflow
[14] provides the API to use portable abstractions or cloud-
si ith service in a workflow application specific features.
di,j data dependency between si and sj
O the union of data dependencies and services
o one element of set O
2.2 Scientific Workflow
ci ith cloud of a set of clouds C A workflow-based application consists of a set of services
Λ possible deployment solutions and data. It is modelled as a Directed Acyclic Graph (DAG),
λ one of the deployments of Λ
Selected set of the services that need to be rescheduled G = (S, E), where S is the set of services, and E is a
Input set of data that have already been generated set of dependencies between those services. Services are
Λ0 possible deployments of the services in Selected represented by the graph vertices and the edges repre-
λ0 one of the possible deployments of Λ0
Cost Model
sent the dependencies between those services. Although a
sci service si is deployed on cloud c workflow-based application can have several different types
Ti,j storage time of data di,j of dependency relationships, in this work we only consider
Storec cost of storing data on cloud c in GB per hour the data dependency (this is the most common dependency
OU T the outgoing data dependencies of a cloud
Comc0 ,c cost of transferring 1GB of data between clouds relationship in scientific workflow applications). In this type
IN the incoming data dependencies of a cloud of dependency, a data item is generated from a source
Tjc execution time of sj on cloud c service and consumed by a destination service. For example,
Execc cost of using compute resources on c for one hour ei,j represents a data dependency between service si and
Scost data storage cost
Ccost communication cost service sj . To represent data dependencies we use a distance
Ecost execution cost matrix D = [di,j ] of size |S| × |S| where a positive value of
Icost initial cost for setting up a new deployment di,j indicates a dependency between si and sj as well as
Dcost cost of the new deployment the size of transmitted data. O represents the union of D
COST total cost of a workflow deployment
Security Model and S . Furthermore, C represents a set of clouds which are
f unc1 embeds constraints for each o available for deployment.
f unc2 represents the constraints for the whole workflow
NW D no-write-down
N RU no-read-up 2.3 General Security Model
SIC security in cloud computing An application’s security can be improved by two ap-
TABLE 1: Notations proaches: firstly, refining the design and implementation
of the application; secondly, deploying the application over
more trustworthy resources, such as shifting the application
to a higher security server. In this paper, we propose to
increase the security of a workflow by adopting the latter
approach. To achieve the enhancement in workflow secu-
rity, we present two functions which are used to provide
a concrete representation for different types of security
requirements. We assume that Λ represents the possible
deployment solutions for given workflow over federated
clouds C and λ is one deployment of Λ, noting Λ = O × C
and λ ∈ Λ.
func1) embeds constraints for d and s. Thus, if λ is a valid
deployment solution, each o ∈ O has its security constraints
and must be deployed on a cloud c ∈ C which can meet the
constraint.
func2) represents the constraints for the whole workflow
deployment. Therefore, a valid deployment solution λ must
meet the security constraint H . Where H is one of the
Fig. 1: Architecture of Cloud Service Broker for Scientific Work- security requirements.
flow
2.4 Cost Model
The cost model is designed to calculate the cost of deploying
the Scheduler, and the Data Manager maintains the data a workflow over a set of available clouds, including data
transfer during workflow execution. The planned tasks are storage cost, data communication cost and computation
distributed to the underlying cloud providers via Deploy- cost. We assume that the clouds are linked in a fully con-
ment APIs. These APIs can also be used to interact with nected topology and the data can be transferred between
the underlying clouds to monitor workflow execution and clouds without obstructions. Additionally, a cloud c can run
cloud availability. several services s at the same time. Therefore, a set of cost
Federated cloud is a cloud resource pool that provisions functions is defined as follows:
computation and storage resources, as well as specific non- The first function is the data storage cost:
functional capabilities (referring to different security levels X
in this paper) and functional capabilities such as the ex- Scost(sci ) = di,j × Ti,j × Storec (1)
ecution environment of each tasks. For example, Jclouds di,j ∈OU T

Where sci means that service si is deployed on cloud c. In the following, we use f unc1 as an example to prove
OU T is a set of data dependencies, representing the data that the optimisation problem is a NP-complete problem.
that are generated by si and transferred to its immediate Theorem: The optimisation is a NP-complete problem.
successor sj which is not deployed on c (note that if all Proof: we first verify that the problem of deploying a
immediate successors of si are on c, then OU T = ∅). di,j workflow over a set of clouds to P meet security requirements
represents the amount of data which is generated by si is a NP problem (noting ∃ λ ∈ (Λ, W), where W repre-
and consumed by sj . Ti,j denotes storage time of data di,j , sents the security requirements).
which is the required time starting from the generation of The NP-completeness of optimising the cost can be illus-
data until the completion of workflow execution. Finally, trated as follows: we start by transforming PARTITION [15]
Storec is the cost of storing 1GB of data for one hour on (one of six core NP-complete problem) to our problem. Let
cloud c. the instance of PARTITION be a finite Set A = (a1 ...am ) and
In this model, we make an assumption that the data a weight w(ai ). We want to have two disjoint subsets A1 and
remains stored only on the source cloud to avoid double- 1 , A2 ⊆ A, where
accounting for the cost. The reason for storing the outputs
A2 ; A
P P A1 ∪ A2 = A and A1 ∩ A2 = ∅, such
that a∈A1 w(a) = a∈A2 w(a).
of a service even after the generated data has been sent to In order to reduce our problem to a PARTITION prob-
another cloud is to handle a failure of the destination cloud. lem, we assume that a workflow has m numbers of O, and
In this case, the stored data provides a way to resume the two clouds are available for deployment. Further, we do
computation on another cloud without the need to restart not consider the security issue, which means any o can
the whole workflow execution. This can be adapted to be deployed over any of the two clouds. Therefore, we
handle the cloud change problem. can have two sets of deployments C1 = (o1 ....om ) and
The second function, Ccost, is used to estimate the C2 = (o1 ....om ) over the two available clouds
communication cost of transferring data between different Regarding our problem, we need to have two disjoint
services. subsets C10 and C20 , where C10 ∪ C20 = O and C10 ∩ C20 = ∅.
X This match the conditions of the PARTITION problem.
Ccost(sci ) = di,j × Comc0 ,c (2)
Furthermore, w(o)
P represents the cost of deploying o onto
di,j ∈IN 0
the cloud, so o∈C10 w(o) is the cost of set C1 . However,
It is the data transferred from the immediate predeces- the PARTITION problem is to find two disjoint sets with the
sors of service sj (denoted as IN ), which are not in the same weight, which has the same complexity as our prob-
same cloud. Comc0 ,c represents the unit cost of transferring lem that is trying to find two
1GB of data from cloud c0 to c. However, if two services P disjoint sets while
P minimising
the total cost, noting min( o∈C10 w(o) + o∈C20 w(o)).
are deployed on the same cloud, the cost is zero, i.e.
∀c0 = c : Comc0 ,c = 0.
Finally, Ecost(sci ) indicates the execution cost of service 2.6 Dynamic Cost Model
si on c. It is defined as:
Dynamic Cloud resources may affect workflow execution.
Ecost(sci ) = Tic × Execc (3) Situations arise when individual nodes may fail during
the execution, or in some extreme cases the whole cloud
Where Tic is the execution time of si on cloud c, and
is unreachable for several hours. As a consequence of
Execc represents the cost of using compute resources on c
their failure, workflow applications may not be executed
for one hour.
Based on the three cost functions, we can formulate the to completion. Furthermore, new clouds, possibly attractive
COST (λ) function to define the total cost of a workflow because they are cheaper or more secure etc., may become
deployment over a set of clouds: available during the execution of workflow applications.
Therefore, to deal with the dynamism of cloud resources,
X we develop a new cost model that dynamically calculates
COST (λ) = (Scost(sci ) + the cost of deploying uncompleted services over the current
sc ∈λ (4)
i available clouds.
Ccost(sci ) + Ecost(sci )) We assume a set Selected is composed of the services
that need to be rescheduled, including unfinished services
2.5 Deployment Optimisation
as well as the services that have been completely processed,
As mentioned previously, we propose a model for opti- and their outputs are the inputs of unfinished services, but
mising monetary cost as well as meeting the security re- the outputs have not been stored because of the failure of
quirements. Therefore, the optimisation problem is to find a the clouds. The details will be illustrated in Section 4.4.
deployment λ ∈ Λ with two constraints: i) the deployment Input is a set of data which have been already generated
λ must meet the security requirements by belonging to from the processed services and required for services in
either f unc1 or f unc2. ii) the value of COST (λ) should Selected, and stored in the available clouds. Based on the
be minimised to obtain deployments with a low cost of definition, we can have the initial cost for setting up a new
execution. We express this problem as: deployment, which is the cost of storing the input data of
minimise (COST (λ)) Selected. It is defined as:
subject to ∀ oc ∈ λ : f unc1(oc ) := true OR X
f unc2(λ) := true Icost(Selected) = di,j × Ti,j × Storec (5)
∃λ ∈ Λ di,j ∈Input

In addition, Λ0 is the set of possible deployments of the

services in Selected over the available clouds C . Conse-
quently, the cost of the new deployment can be defined as:

Dcost(λ0 ) = COST (λ0 ) + Icost(Selected) (6)

Where λ0 ∈ Λ0 represents one of the possible deploy-

ments for the services in Selected.

3 S ECURE DEPLOYMENT
In this section, we apply our previous work [16], based on
the Bell-LaPadula [17] Multi-Level Security model [18], to
demonstrate how to adapt the security model to DoFCF.
This incorporates the security levels of the clouds, data and
services to achieve a secure deployment for the workflow
over a federated cloud. Fig. 2: The Architecture of DoFCF
In our security model, each service S has two security
levels: “Clearance” and “Location”. “Clearance” represents
the services’ highest security level, and “Location” is the have proved our problem is a NP-Complete problem, there-
required operation security level of the service in a specific fore it is very difficult to design an algorithm using B&B in
application. The data D and cloud C only have “Location”. polynomial time as a generic framework.
l(o) and c(o) represent the security of location of o and the Although this method gives the optimal solution and
clearance of o respectively. guarantees that the result is the cheapest deployment, it is
W represents the security constants, including three not very scalable. In our paper [20], we demonstrated that
rules: when the number of services increased to 12, a version of
B&B that we have implemented, required approximately 15
• NWD “no-write-down”: denotes that a service can- minutes to generate a solution. Thus, this type of algorithms
not write data which has a lower security level are not considered in our framework.
(required security level) than its own. This can be for-
malised as: N W D(di,j , sj ) = c(sj ) ≥ l(di,j ) ? true :
4.2 Genetic Algorithm
f alse
• NRU “no-read-up”: means a service cannot read A Genetic Algorithm (GA) can efficiently find a solution
data if the data’s location security is higher than the to a problem in a large space of candidate solutions [21].
service’s clearance security, noting: N RU (si , di,j ) = It is a search heuristic that mimics the process of natural
l(di,j ) ≥ l(si ) ? true : f alse selection to find the optimal solution, yet in our case the
• SIC “security in cloud computing” (SIC): defines heuristic function will not constantly produce the optimal
the location security level of a cloud that should be (or cheapest) solution. Moreover, the design or method of
greater than or equal to the location security level application of GA can also have a significant impact on the
of any service or data that are hosted on this cloud– quality of the solution [22].
SIC(dci,j , scj ) = l(c) ≥ l(si )&l(c) ≥ l(di,j ) ? true : In the following, we extend and adapt GA to our frame-
f alse. Where dci,j and scj represent data di,j and work to find an acceptable solution in polynomial time.
service si to be deployed on cloud c.
4.2.1 Security Candidate List
In APPENDIX, available in the online supplemental In this paper, we are aiming to find an optimised solu-
material, we run a deployment example to show how to tion that meets the security requirements while minimising
apply the above security rules to a workflow application. the monetary cost. Therefore, this can be considered a bi-
objective optimisation problem. As mentioned in Section 3,
4 D EPLOYMENT O PTIMISATION A LGORITHMS each object of the workflow has its security requirements for
deploying over clouds. Therefore, we firstly list the satisfied
In this section, we investigate and analyse some state-of-
clouds for each object of the workflow in “Candidate List”.
the-art optimisation algorithms and then extend the Genetic
The security requirements for each object can be hard
Algorithm (named adaGA) and Greedy Algorithm (named
constraints (noting it must be met), and the valid clouds that
NCF) to adapt to our framework. The architecture of our
meet these constraints will be maintained in the “Candidate
new framework DoFCF is depicted in Figure 2.
List”. Consequently, our problem is reduced to a single
objective optimisation which is minimising the monetary
4.1 Branch and Bound Algorithm cost of the deployment.
As discussed above, we need to find a λ ∈ Λ which min-
imises the deployment cost. The most common approach is 4.2.2 Elitist Prevention and Diversity Maintenance
B&B (branch and bound algorithm) [19]. Generally, this type The basic GA can be adapted to generate a deployment solu-
of algorithms require ranking all of the secure deployment tion for the problem discussed above. However, to generate
solutions and then choose the cheapest one. However, we an efficient solution, two primary factors: selection pressure

ALGORITHM 1: Elitist Prevention

s–the size of elitist list; elist–list of elitist individuals; pop–list of
all individuals
if elist is empty then
ASCsort (pop) . ASCsort sort the pop as ascending order
elist ← from pop[0] to pop[s − 1]
. copy the first s number of solutions to elist
end
for o in O do Fig. 3: The overview of adaGA
for c in C do
if o is securely deployed on c then
SArray[o] ← c
else 4.2.3 Adapting to the Framework
pop ← combine elist and pop The adaption of a genetic algorithm has been divided into
ASCsort (pop)
delete s numbers of pop in tail five phases: coding, generating a candidate list, initialising
elist ← from pop[0] to pop[s − 1] individuals, selection, crossover and mutation. We coded
end j j
our deployment solution as a vector [si1 , s2 ....skn ], where si
end
end means that service si is deployed on cloud cj . In order to
reduce the possibility of generating an insecure solution, we
chose the clouds from “Candidate List”, assigned them to
the corresponding objects, and then coded them as above.
ALGORITHM 2: Diversity Protection However, applying these operations will not avoid produc-
pop–a list of all individuals; size–size of pop; ing a few new coded solutions which may not meet the
threshold–predefined threshold of diversity; rate–the current security requirements. In such cases, rule SIC is used to
mutation rate; M inrate–minimum mutation rate; verify the security of those solutions. If insecure code is
M axrate–maximum mutation rate.
. function removeDup removes the duplications detected, the algorithm will generate a new one to replace
of pop and then copy it to rpop it.
rpop ← removeDup(pop) After generating the initialised individuals, selection,
sr ← size of rpop
sr
d ← 1 − size
crossover and mutation operations are applied on the in-
if d > threshold and rate < M axrate then dividuals to generate new generations. During this process,
increase rate Algorithm 1 is used to prevent elitist individuals, and two
end methods are applied to perform the selection: one is a fitness
else if rate > M inrate then
decrease rate function that can transfer the fitness of a coding into a nu-
end meric representation to select superior solutions. For this we
can use equation 4 above. The other is a diversity analysis
that does not impact the selection result, but influences the
crossover and mutation rate.
and population diversity have to be considered carefully. Crossover then takes part of features from two chromo-
Moreover, these two factors are inversely related, and so the somes and combines them to generate a new chromosome.
GA must be carefully designed to balance the effect on the The crossover we used is one-point crossover [24].
population of diversity and selection pressure. In order to To enhance the search range, mutation is used. This is
solve this trade-off, we apply the elitism method [23] sum- implemented by randomly selecting the chromosomes in
marised in Algorithm 1, to increase the selection pressure, the current generation and changing them to new secure
and at the same time control the diversity dynamically. chromosomes. This operation only happens with very low
probability as described in [25]. The overview of adaGA
The purpose of the elitist method is to avoid destroy- with Elitist Prevent and Diversity Protection Algorithms is
ing superior individuals in crossover and mutation. Thus, shown in Figure 3, the whole processes are repeated until
once a solution is confirmed as elitist, it should be directly the termination constraints are researched.
inherited by the new generation of the population. {In this paper we do not consider how to optimise the param-
Low diversity of a population usually means that the eter setting due to the limited space, however in [26] and [25]
search reaches a local extrema, which significantly impacts authors have explored more justification on parameter setting,
the solution. In order to solve this problem, we dynamically such as population size, crossover and mutation probability and
control the mutation rate to influence the generation of the stop criteria.Therefore, in our experiment we just follow their
new chromosome. Algorithm 2 shows how the diversity techniques about how to set the parameters of GA.}
protection works. In Section 5 the evaluation of our algorithm will be
presented and compared with HUGA (Hybrid Utility-based
Firstly, we remove duplications in pop, and count the
Genetic Algorithm) [27] which was used to optimise the de-
total number (sr) of unique solutions. Based on that we can
ployment of workflow over a federated cloud, considering
have density d of the population, which is represented by
QoS requirements.
the ratio of duplicated solutions. Next, the mutation rate
will be increased if the density is greater than the predefined
diversity threshold and the mutation rate is less than the 4.3 Greedy Algorithm
maximum mutation rate. If the mutation rate is greater than The greedy algorithm is an iterative algorithm that incre-
the maximum mutation rate, it will be decreased. mentally finds better solutions. Unlike the Genetic algo-

rithms that need to finish executing before returning a solu- ALGORITHM 3: NCF
tion, the greedy algorithm generates a valid and improved W –workflow; S set of service; D–set of dependencies between
solution in each iteration. This is a desirable characteristic related services; C –set of available clouds
if not((WorkflowSecurity(D,S ))) then
for systems where the parameters change frequently and Invalid Workflow
the available time for calculating an improved deployment else
varies significantly. INI=InitialDeployment(D,S ,C )
end
In this section, we develop a method for finding a de- function WorkflowSecurity(D,S )
ployment solution as an extension of the NCF (Not Cheapest for di,j ∈ D do
First) [20] algorithm to adapt it to our framework. The NCF if not (N RU (di,j , sj ) and N W D(si , di,j )) then
is an extended version of greedy algorithm, which uses extra return False
end
information for planning a deployment, making short term end
sacrifice for long term benefit. NCF, as summarised in Algo- return True
rithms 3 and 4, pre-deploys each service on the cloud which function InitialDeployment(D,C )
. Topsort returns a topological order of W
minimises the cost and meets the security requirements in for si ∈ topsort(W) do
isolation, therefore applying a set of optimisation methods for cj ∈ C do
to refine the pre-deployment. if SIC == T rue then
c
The algorithm consists of the following three steps. First, if T emp[si ] > COD(si j ) then
c
it starts by applying security rules to verify whether security T emp[si ] ← COD(si j )
end
requirements are met by the original workflow. The work- end
flow is valid iff all return values from NRU and NWD are end
true. Otherwise, the workflow is invalid, the security check end
return T emp
returns an error and the whole algorithm stops.
Next, we calculate the cost of deploying services on each
valid cloud, using the COD function. COD is calculated by
ALGORITHM 4: NCF (Refinement)
adding the computing cost of service si to the transmission
si,max – child service of the service si with maximum COD
cost and storage cost of data sent from all its immediate pre- value; SET P (si )– a set which includes service si,max and all its
decessor services that are not in the same cloud. The initial parent services; SET C(si )– a set includes si and all its child
deployment of services is based on the smallest COD value services.
of each service taking into account the security requirements U S = topsort(W)
for si ∈ U S) do
checked by SIC rule. Function InitialDeployment runs switch(Psi )
until each service finds a cloud that can meet the security case1: sh ∈SET P (si )
IN I(sh ) > M IN (SET P (si )) then
requirement and its smallest COD value associated with this Deploy all services in SET P (si ) to the cloud which
cloud is stored in vector T emp. minimises
P the cost and remove(SET P (si )) from U S .
case2: sh ∈SET C(si )
IN I(sh ) > M IN (SET C(si )) then
Deploy all services in SET C(si ) to the cloud which
COD(sci ) =Scost(sci ) + Ccost(sci ) minimises the cost and remove(SET C(si )) from U S .
+ Ecost(spi ) case3: both case1 and case2 are satisfied then
if M IN (SET P (si ) > M IN (SET C(si ) then
Finally, the core idea behind function Ref inement is to Deploy all services in SET C(si ) to the cloud which
avoid scheduling services to clouds with huge communica- minimises the cost remove(SET C(si )) from U S .
else
tion costs. This function includes four cases which detect the Deploy all services in SET P (si ) to the cloud which
services of initial deployment that can be refined. Since the minimises the cost remove(SET P (si )) from U S .
services have been found, these services are assigned to a end
case4: both case1 and case2 are not satisfied then
cloud which minimises deployment costs, while this cloud Deploy si to the cloud which minimises its COD value and
must meet security rule SIC. si remove U S
end

4.4 Adaptive Rescheduling Algorithm

In this section, we introduce a heuristic algorithm which
has adapted to the DoFCF framework and dynamically monitored until the workflow has been fully executed.
generated a new solution for handling cloud availability Alongside this, a new deployment must be generated in
changes. polynomial time, because the time is accounted for in the
The generic adaptive rescheduling algorithm for han- makespan of the workflow execution, which may bring
dling the change of cloud availability works as follows: extra cost. In extreme cases, if it takes considerable time,
when a change in cloud availability is detected, the planner the available clouds may change again.
estimates the monetary cost for each service based on the We designed and developed a dynamic rescheduling
available clouds and information (e.g. execution time) of algorithm to be adapted to our framework, which is sum-
each service. For our case, the estimation heuristic algorithm marised in Algorithms 5 and 6. In this algorithm, we
can be applied to generate a new deployment solution for assume that a workflow W is executed with the deploy-
the services in Selected over available clouds to meet the ment DEP0 . The ”Cloud monitor” is used to monitor
security requirements as well as minimising the monetary the workflow execution status and cloud status. Once the
cost. Therefore, services are distributed based on the new monitor has detected any changes in cloud status, function
solution and then executed. In addition, the execution is Redeployservices will be invoked to find the services which

ALGORITHM 5: Dynamic rescheduling ALGORITHM 6: Redeploy Services

U P – set of unprocessed services; C –set of all available clouds; DEP0 –the original deployment; Selected-set of services for
DEP0 –deployment redeploying; parents(si )–set of si ’s parent services
set initial U P = all services of W function Redeployservices(U P ,C )
while W not finished do Selected ← U P for si ∈ U P do
update C . C is updated via the communication with F IN DN ODES(parents(si ), C, Selected)
CloudM onitor end
update U P . U P is updated via removing the finished return Selected
services function Findnodes(parents(si ),C ,Selected)
if cloud status has changed in C then if parents(si ) == ∅ then
Selected=Redeployservices(U P, C) return
DEP1 = N CF (Selected, C) if clouds have failed in C else
then for sj ∈ parents(si ) do
. submit current execution information to generate if sj 6∈ U P then
a new deployment. . DEP0 (sj ) indicates the cloud on which sj was
if DEP1 is not found then deployed
break; if DEP0 (sj ) ∈ C then
. not valid deployment else
else add sj to Selected if not included
submit DEP1 F IN DN ODES(parents(sj ), C, Selected)
DEP0 ← DEP1 end
end end
. new cloud resources become available end
else
. if the new deployment is cheaper than the State end
present one
if DEP0 .cost > DEP1 .cost then
submit DEP1
DEP0 ← DEP1 phases, we need to operate on the whole chromosome,
else so the complexity of both is O(|P | × |G|). Thus the time
do nothing
end complexity of HUGA is O(|P | × |G| × |O|).
end Comparing adaGA with HuGA, adaGA contains an ad-
end ditional Elitist phase for each generation which slightly
end
increases the time complexity. In each Elitist phase, the
algorithm observes the chromosomes for each population
and saves the best one. Thus, the complexity of this phase is
require redeployment. These actions are summarised in Al- O(|P | × |G| × |O|). Therefore, the time complexity of adaGa
gorithm 6 as follows. Firstly, the availability of input data for is O(|P | × |G| × |O|) as well.
services in set U P is checked, which means that all inputs In the NCF algorithm, we have already analysed the
data must be stored in the available clouds. Otherwise, the time complexity in paper [20]. It is significantly impact by
elements in set U P will be copied to the Selected set and the structure of the workflow. if the workflow is linear, the
added to the services which have been completely executed complexity in the worst case becomes O(|O| × |C|). Con-
but their outputs are not available to Selected. Then, the versely, for a star-shaped workflow the best case complexity
function tracks back and repeats the same action until the is O(|D| × |C|).
available inputs are found. In this case, Selected will include
the services in U P as well as the added services. Otherwise,
Selected is equal to U P . 5 E XPERIMENTS AND E VALUATION
Based on the selected services and the available clouds, To evaluate the performance of our proposed framework,
a new deployment DEP1 can be generated by applying a we conducted a series of simulation experiments on a num-
NCF algorithm. If the cloud change is caused by the failure ber of real scientific workflow applications. In addition, we
of some clouds, DEP1 is used to deploy the services in applied our framework to dynamically deploy a scientific
Selected directly. Otherwise, if the change is caused by a workflow over a set of e-SC instances on multiple clouds.
new cloud becoming available, then the cost of DEP1 and The experimental setting, and results are presented in the
DEP0 will be compared (only considering the unfinished following subsections.
services). If DEP1 is cheaper, the workflow will be sched-
uled to a new deployment solution, otherwise the execution 5.1 Simulation Environment
status and cloud status are monitored until the workflow is
In this work, the experiments must be repeatable in order
completely executed.
to easily compare and analyse different types of algorithms.
Therefore, to ensure repeatability, we evaluated our frame-
4.5 Time Complexity work using CloudSim to investigate the deployment and
The time complexity of each algorithm is formalised as scheduling of workflows in federated cloud environments.
follows:
HUGA algorithm can be split into three phases in order 5.1.1 Experiment Setup
to analyse the time complexity: selection, crossover and mu- In the evaluation of our framework, we consider four com-
tation. Therefore, the time complexity of the selection phase mon types of workflow applications: CyberShake (earth-
is O(|P | × |G| × |O|) where P is the size of population and quake risk characterisation), Montage (generation of sky
G is the number of generations. For crossover and mutation mosaics), LIGO (detection of gravitational waves) and

2168-7161 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2016.2603477, IEEE
Transactions on Cloud Computing
JOURNAL OF LATEX CLASS FILES, VOL. ??, NO. ??, ??? 9
Workflow Medium Large Very large
CyberShake 30 100 1000
The results show that the algorithm adaGA can always
Montage 25 100 1000 generate the cheapest deployment solution. For example,
LIGO 30 100 1000 in a case with medium size of “Montage” workflow, the
Epigenomics 24 100 995 solution generated by adaGA can save up to 35% compared
to the NCF solution.
TABLE 2: Number of Tasks of each Workflow for Each of the
Three Scales The types of workflow significantly impacted the solu-
tions generated by NCF. {Fig 4 shows that the costs of the
Type Location Exec Store In Out deployment solutions generated by the three algorithms are very
(/hour) (/hour/GB) (/GB) (/GB) close when these algorithm are applied to the workflows of the
C1 0 0.40 0.10 0 0.02
C2 2 2.20 0.60 0.03 0.01
LIGO and Epigenomics in Medium size. } However, for the
C3 1 1.23 0.30 0.14 0.07 other two types of workflows, the solutions generated by
C4 2 3.70 0.60 0.10 0.05 NCF are much more costly than adaGA. Furthermore, the
C5 3 4.50 0.90 0.14 0.05 differences are reduced with the increase in workflow size.
C6 4 5.5 1.30 0.14 0.13
This is because NCF is not influenced by the search space
TABLE 3: Cloud Pricing and Security Levels (larger workflow indicates more deployment solutions). In
addition, the search space significantly impacted the results
generated by adaGA and HUGA. However, the Elitist Pre-
Epigenomics (bioinformatics).1 The full characterisation of vent and Diversity Protection methods were used in adaGA
these workflow applications can be found in [28], however, to avoid the algorithm visiting the less desirable solutions.
we only consider the execution time, and the input and
output data of each service. Table 2 lists the four workflow 5.1.3 Time Complexity Evaluation
types with different numbers of tasks: medium, large and In order to evaluate the time complexity of each algorithm,
very large. we measured the time consumed by each to find the op-
The data privacy information for the workflows is un- timised deployment for the four types of workflow on the
available to be used for assigning the security levels of each given clouds. According to our evaluation, Fig 8 shows that
object. Therefore, we randomly generated the security levels algorithm NCF is significantly faster than the other algo-
for each object in these workflows. rithms. Further, adaGA has better performance than HUGA
Six VMs have been created, representing workflow exe- with medium size workflows. The reason is that the search
cution environments, in six different data centres to repre- will be terminated if no better solution has been found
sent six types of cloud (with different security levels). Addi- after repeating the pre-defined generations. Nevertheless,
tionally, each VM can run several services at the same time. as the workflow size increases, adaGA consumes more time
In this paper, we do not consider the performance issue. than HUGA to find the deployment solution. However, the
Therefore each cloud shares the same configuration, within deployment solution for a very large size of “CyberShake”
1 core, 2 GB RAM and 12GB Disk. Table 3 details number workflow can be generated by adaGA in less than one
of clouds with location security levels, computation cost, minute. Consequently, by considering cost savings, adaGA
storage cost and communications cost, where In and Out can be the better choice.
represent the transferring cost of incoming and outgoing Fig 8 also indicates that the time complexity of each
data respectively. algorithm is not only dependent on the number of o (data
The experiment results, presented below, are the average and services) and the available clouds, but also on the
values of observing 1000 executions of each algorithm for structure of the target workflows and the security levels for
each type of workflow. For each of the 1000 repetitions, the each o and cloud.
same random number generation seeds for each execution,
which guarantees each algorithm is exactly running over the 5.1.4 Cloud Availability Change Evaluation
same infrastructure. In this part of the experiments, we simulated the change in
cloud availability by predefining the times when each cloud
5.1.2 Monetary Cost Evaluation was available. To do this we set the start time and termi-
As discussed earlier, the total cost includes execution cost, nation time for each cloud before starting the simulation. A
running time storage cost and communication cost. The Cloud Monitor was implemented to monitor cloud status, i.e.
pricing of each cloud listed in Table 3 shows that a more detect changes in cloud environments. If a changed status
secure cloud is generally more expensive. Our cost calcula- is detected, a notification is sent to the broker. Thus, the
tion does not consider additional costs like license and VM broker can reschedule the running workflow to the available
image costs which are charged by some cloud providers. clouds based on the cost of the new deployment. This was
Figs 4, 5 and 6 depict the normalised cost for the evaluated for two types of change: Clouds fail and Availability
four types and three size of workflow applications men- of new clouds.
tioned previously using three algorithms: NCF, HUGA, and Cloud fail was simulated by setting the terminal time
adaGA. Each figure represents cost calculations for a specific for the randomly selected clouds as uniformly distributed
workflow size to show the variations in the cost according between 0 and the makespan. In another words, during
to the size. the workflow execution, the number of cloud failures is
randomly between 0 to 6. Furthermore, we performed 1000
1. The XML description files of the workflows are available via the
Pegasus project: https://confluence.pegasus.isi.edu/display/pegasus/ simulations, each with different cloud failure settings, and
WorkflowGenerator recorded each execution, including how many clouds fail,

the assigned tasks before their terminating time. (2) SBR:

the workflow is successfully executed by rescheduling on
the currently available clouds after a cloud has become
unavailable. (3) Fail: the workflow execution cannot be com-
pleted as no alternative deployments were available after a
cloud failure. This would be because of cloud unavailability
or because the available clouds do not meet the security
requirements.
Table 4 depicts the results of cloud failure. These demon-
strate that the failures occur more frequently with the
increasing sizes of workflow and execution time. We nor-
malise the results by using the ratio of value of the outputs
Fig. 4: Cost for Medium Size Workflows of the SBR executions with that from the Success executions,
recording SBR/Success. In the case of a very large work-
flow, we only have to pay an extra 2% money and 33% time
to avoid re-running the workflow from the beginning when
a failure happens during the workflow execution.
Workflows Execution status (%) Cost Time
Success:74 1 1
SBR:16 1.14 2.36
Medium
Fail:10 none none
Success: 70 1 1
Large SBR: 20 1.1 1.15
Fail:10 none none
Success: 54 1 1
Very Large SBR: 22 1.02 1.33
Fail: 24 none none
Fig. 5: Cost for Large Size Workflows
TABLE 4: Experiment results for cloud failure

To simulate Availability of new clouds, the workflow has to

be already running over a set of clouds with a deployment.
Thus, we pre-set clouds C6 and C5 as available for the initial
deployment. Furthermore, we randomly selected clouds
and set the start time of them as uniformly distributed over
[0, ∞]. We generated 1000 settings and repeated the execu-
tions of the rescheduled workflow to new clouds based on
the monetary cost savings. This resulted in the execution
types shown in Table 5: (1) Success represents no new
clouds becoming available or the new clouds is not offering
cheaper deployment than the currently running one. (2) SBR
denotes the running workflow can be rescheduled to the
Fig. 6: Cost for Very Large Size Workflows
new available clouds to save execution costs.
Fig. 7: The observed outputs metrics are execution cost nor- Table 5 shows that the fluctuations of the saving cost are
malised with the correspondent value obtained from the NCF very significant, depending on the types of the workflow
algorithm, noting Others/N CF and available clouds. (In Table 5, we used the same normal-
isation rule as Table 4). Therefore, when new clouds become
available, users should participate in making the decision to
the makespan, completion of workflow execution. Con- use a new deployment based on the estimated saving cost
sequently, we can have the average of the cost and the and makespan.
makespan based on the recording. Moreover, Table 5 shows that the makespan goes up,
In this evaluation, we used “Epigenomic ” workflow while the monetary cost goes down. The reason is that when
with medium, large and very large size. As an example a running task is shifted to a new cloud which is cheaper, the
of the setting for the medium workflow size, we randomly running task has to be killed, and then be redeployed and
selected the clouds and set their termination time as uni- re-executed in the new cloud. If the new cloud is faster than
formly distributed over [0, makespan] where the makespan the current host cloud, the makespan might be reduced.
is the execution time of the selected workflow running as the However, in this work we do not consider the performance
deployment generated by NCF. For the clouds that were not of the clouds and assume that clouds’ performance are the
selected, the termination time was set as infinite. same.
Thus, we can have three types of executions: (1) Success: 5.2 Realistic System Evaluation
in this type, the workflow is completely executed without To evaluate our algorithm in conditions closer to a produc-
rescheduling as the selected clouds finished executing in tion use, we applied it to schedule scientific workflows in

10
NCF adaGA HUGA NCF adaGA HUGA
30
8

6 20

Ratio

Ratio
4
10
2

0 0
CyberShake LIGO Montage Epigenomics CyberShake LIGO Montage Epigenomics
Medium Size Workf ows Large Size Workf ows

NCF adaGA HUGA

200

150
Ratio

100

0
CyberShake LIGO Montage Epigenomics
Very Large Size Workf ows

Fig. 8: The observed outputs metrics are execution time normalised with the correspondent value obtained from the NCF
algorithm, noting Others/N CF

Workflows Execution status (%) Cost Time shown in Tables 6 and 7. Note that the data size transferred
Success:92 1 1
Medium SBR:8 0.98 1.46
among blocks and the execution time of each block are real
Success: 88 1 1 values taken from logs collected by e-SC. Table 6 shows data
Large SBR: 12 0.90 1.21 sizes in GB, where 0 denotes less than 1 MB of data. The
Success: 70 1 1 pricing of Clouds C1, C2 and C3 in Table 3 was applied to
Very Large SBR: 30 0.99 1.33 calculate the deployment cost.
To simulate this environment we set up three virtual
TABLE 5: Experiment results for new cloud available
machines, each running a single instance of e-SC system.
VM1 was hosted on a personal PC and represented the
private cloud. Two other VMs were hosted in our University
e-Science Central(e-SC). We used the e-SC APIs to create a
virtualised environment and played the role of public cloud
Cloud Services Broker that can orchestrate invocations of a
providers C2 and C3.
single workflow partition over a number of e-SC instances.
Service Name Clearance Location Time(/h)
5.2.1 Design and Setup Sample N ame S1 1 0 1
ImportF ile S2 1 0 1.5
According to the architecture of the cloud services broker, Sample V alue Info S3 1 0 3
shown in Fig 1, our tool consists of three components: Client, HG19 S4 1 0 0.1
Cloud Services Broker and Federated Cloud. F ilter S5 2 0 10
Exome − Regions S6 1 0 7
The Client includes a user interface (UI) which allows Intervalpadding S7 0 0 20
users to create workflows for the e-Science Central work- ColumnJoin S8 2 0 0.1
flow engine. The description of the created workflow can AnnotateSample S9 2 0 5
Export S10 1 0 0.3
then be passed to the Broker. Cloud Services Broker is the core
part of our tool and includes a planner to assign workflow TABLE 6: Services representation and security and execution
to federated cloud using the algorithms discussed earlier. e- time
SC APIs are used to dispatch tasks to corresponding clouds
and monitor the execution. Failure Generator is used for
simulating failures by turning on or shutting down e-SC Data Location Size (GB)
S1,8 1 0
instances. Federated Cloud is a set of e-SC instances which can S2,5 0 1.1
interact with the broker and other e-SC instances through e- S3,8 2 0.01
SC APIs, and process the tasks which are scheduled. S4,5 0 0.005
S4,7 0 0.005
To evaluate our tool we selected one of the workflows S5,7 0 6.2
used in the cloud e-Genome project [29]. S6,7 0 10.3
The workflow was implemented to process exome se- S7,9 1 3.6
quenced by using e-SC deployed on Microsoft Azure cloud. S8,9 0 0
S9,10 0 0.05
While in the e-Genome project security aspects are not
a primary concern, guaranteeing that human genomic data TABLE 7: Data security and size
can be securely processed on the cloud is very important.
Therefore, we modelled the security requirements of the Our evaluations included three steps: the first step tests
selected e-Genome workflow by assigning security levels as the static deployment algorithm. We kept all three e-SC

instances running and applied adaGA to make the deploy- inputs). It is because C1 has to re-execute S2 and S6 from
ment plan. The second step shows how to handle a cloud the beginning.
failing, by shutting down one of the running e-SC instances
when the workflow was running. The setting of this step Deployment Time (seconds) Cost
Static 28.93 64.44
is similar to that of CloudSim. Finally, we tested the avail- SBRF 28.94 67.23
ability of a new cloud by deploying the given workflow on ININ 28.92 84.62
two clouds, and then turning on a new instance which offers SBRN 29.90 65.17
price advantage. Also, for the purpose of the experiment we
TABLE 9: The cost of different deployments
reduced the execution time of the given workflow to about
30 seconds by scaling down the amount of input data shown
in Table 7 by a factor of 6000. 6 R ELATED W ORK
Cloud computing is a technology for transforming com-
5.2.2 Results and Analysis puting as utility model such as water, electricity, gas and
Based on the presented experiment setup, all of the three telephony [30]. Therefore, it is unlike grid in that the total
steps of the deployments are illustrated in Table 8. Precisely, ownership cost of running a workflow is considered to be
“Static” refers to first step, and “Cloud fail” and “New a much more important optimisation criterion. Compared
Cloud” correspond to steps two and three respectively. with other computing resources, cloud has unique features:
Static, shown in Table 8, represents the cheapest solution pay-as-you-go pricing, multiple instance types, elasticity,
which was generated using the adaGA algorithm by de- without operation of infrastructure and so on. Thus, the
ploying the workflow over cloud C1, C2 and C3 to meet the state-of-the-art techniques or mechanisms for workflow
security requirements. Services S1 S7 and S8 were deployed management need to be adapted to the new computing
on C2 and others were allocated on C1. environments.
For the Cloud Fail, we used the deployment of Static On a single cloud, most research efforts are aimed at
as the initial deployment (INI). However, cloud C1 failed improving the performance of workflow systems. In [31],
(shown as blue in Table 8) when service S9 was ready to the authors introduced an auto-scaling method that applied
execute. a fixed sequence of transformations on the workflow in a
The available clouds are C2 and C3, and the inputs of S9 heuristic way. However, the sequence of workflow transfor-
are stored in C2 (the outputs of S7 and S8 ), therefore, S9 can mations are not unique, and different transformations have
be rescheduled to C3 to continue the execution (indicated in quite different costs.
“Cloud fail”, SBR). If S7 is deployed on C1 with the same The most common approach is concentrated on work-
failure, S2 , S4 , S5 , S6 and S7 should be re-executed in C3. flow scheduling for running workflow in cloud to meet
Thus S9 and S10 can be completed in C3. performance and cost constraints. The authors in [32] de-
In the third step, C2 and C3 were available for the scribed a method which can dynamically provision VMs
initial deployment (see Table 8 New Cloud, INI). After for meeting the performance requirements of executing
the workflow was executed for one second , C1 became workflows, and recover the computing resources when they
available. S1 and S4 were completely executed on C2 and are over provisioned to reduce the monetary cost. Kllapi et
C3 respectively, but S2 , S6 and S3 were still running. Based al. presented a method to deal with the trade-off between
on the status information, a cheaper deployment solution makespan and financial cost for data processing flows [33].
(see New Cloud SBR in Table 8) became available, which The authors in [11] proposed an algorithm that uses the
required termination of S2 and S6 , and then re-running idle time of provisioned resources and surplus budget to
them on C1, as shown in green in Table 8. replicate tasks so as to increase the likelihood of meeting
deadlines. At the same time, the economic cost of execution
Service Static Cloud fail New Cloud is also minimised by carefully planning the provision of
INI SBR INI SBR VMs.
S1 C2 C2 C2 C2 C2
S2 C1 C1 C1 C3 C1 Considering security-driven scheduling, only few
S3 C2 C2 C2 C2 C2 groups of researchers have investigated this topic from
S4 C1 C1 C1 C3 C3 different angles in various contexts. Mace et al. [34] explored
S5 C1 C1 C1 C3 C1
S6 C1 C1 C1 C3 C1
the current information security issue of public cloud and
S7 C2 C2 C2 C2 C2 provided general security solutions to choose what work-
S8 C2 C2 C2 C2 C2 flows or subsets of workflows can be executed in a public
S9 C1 C1 C3 C3 C1 cloud while ensuring the security requirements are met. The
S10 C1 C1 C3 C3 C1
authors in [35] proposed a security-driven scheduling algo-
TABLE 8: Two deployments rithm for DAGs workflow which can achieve high quality of
application security, based on task priority rank to estimate
Table 9 shows average values of the cost and makespan the security overhead of tasks. In addition, this algorithm
of each deployment by repeating the executions 10 times. considered improving the performance of workflow execu-
Where SBRF represents the situation of handling C1 fail tion.
(see Table 8 Cloud Fail SBR). Similarly, ININ and SBRN are Cloud federation work is relatively new in the cloud
the experimental results of new cloud available, where the computing area. Therefore, there is little available literature.
makespan of SBRN is approximate one second more than In [36], the private cloud (the user’s own machines) was
others (it will take one hour more by using the original also assumed to be a free computing resource, with limited

computing power. A public cloud such as Amazon EC2 can consider the risks of cloud computing, but none of them
meet users’ performance requirements, but the cost must provides a quantitative measurement. This measurement
also be minimised. A framework, called PANDA (PAreto can be a strong support for this paper by providing the
Near optimal Deterministic Approximation), was designed realistic security level of each cloud datacenter.
for scheduling workflow across both public and private
clouds with the best trade-off between performance; hence
Pareto-optimality. Fard et al in [37] introduced a pricing ACKNOWLEDGMENTS
model and truthful mechanism for scheduling workflow to This work was supported by the RCUK Digital Economy
different clouds, considering the monetary cost and comple- Theme [grant number EP/G066019/1 - SIDE: Social In-
tion time. In order to solve the trade-off between cost and clusion through the Digital Economy], and the European
performance, a Pareto-optimal solution is adapted in the Communitys Seventh Framework Programme (FP7/2007-
scheduling algorithm. However, none of them considered 2013) under grant agreement n. 600854 SmartSociety Hy-
security and cloud availability change. brid and Diversity-Aware Collective Adaptive Systems:
SABA (Security-Aware and Budget-Aware workflow Where people meet machines to build smarter societies
scheduling strategy) [9] provided a static workflow deploy- (http://www.smart-society-project.eu/).
ment solution over multi-clouds for optimising security,
makespan and monetary cost. The optimisation in this work
was based on a heuristic list which ranks the priority of R EFERENCES
each task of the workflow through a normalisation function.
[1] G. Juve and E. Deelman, “Scientific workflows in the cloud,” in
Jrad et al. [27] proposed a cloud broker to schedule larger Grids, Clouds and Virtualization. Springer, 2011, pp. 71–91.
scientific workflows over federated clouds to match the QoS [2] M. Armbrust, A. Fox, R. Griffith, A. D. Joseph, R. Katz, A. Konwin-
and cost requirements. Since these two efforts are static ski, G. Lee, D. A. Patterson, A. Rabkin, I. Stoica, and M. Zaharia.
(2009) Above the clouds: A berkeley view of cloud computing.
scheduling algorithms, they are unable to manage the cloud [3] F. Zhang and M. Sakr, “Performance variations in resource scaling
availability change. for mapreduce applications on private and public clouds,” in
Regarding exception handling, the data flow-based ap- Cloud Computing (CLOUD), 2014 IEEE 7th International Conference
proach in [38] was introduced to support hierarchical ex- on, June 2014, pp. 456–465.
[4] S. Chaisiri, B.-S. Lee, and D. Niyato, “Optimization of resource
ception propagation and user-defined exception. This work provisioning cost in cloud computing,” Services Computing, IEEE
considered the exceptions caused by workflow itself, while Transactions on, vol. 5, no. 2, pp. 164–177, April 2012.
we focused on solving computing resource change prob- [5] Z. Zheng, H. Ma, M. Lyu, and I. King, “Collaborative web service
qos prediction via neighborhood integrated matrix factorization,”
lems.
Services Computing, IEEE Transactions on, vol. 6, no. 3, pp. 289–299,
Furthermore, in our previous work [10], we proposed July 2013.
a dynamic method to handle cloud failure issues, while [6] K. Finley. Godaddy outage takes down millions of sites,
the workflow is running on a federated cloud. However, anonymous member claims responsibility, year = 2012, url
= http://techcrunch.com/2012/09/10/godaddy-outage-takes-
this method can not handle large scale workflow and also down-millions- of-sites/, urldate = 2012.
did not consider the situation that a new cloud becomes [7] (2011) Summary of the amazon ec2 and amazon rds
available. service disruption in the us east region. [Online]. Available:
http://aws.amazon.com/message/65648/
[8] R. Buyya, R. Ranjan, and R. N. Calheiros, “Intercloud: Utility-
oriented federation of cloud computing environments for scaling
7 C ONCLUSION AND F UTURE W ORK of application services,” in Proceedings of the 10th International
In this paper, we have presented the DoFCF framework to Conference on Algorithms and Architectures for Parallel Processing -
Volume Part I, ser. ICA3PP’10. Berlin, Heidelberg: Springer-Verlag,
improve the security of the workflow applications while
2010, pp. 13–31.
they are distributed on a federated cloud. A cost model [9] L. Zeng, B. Veeravalli, and X. Li, “Saba: A security-aware and
has been designed to optimise the cost of each deploy- budget-aware workflow scheduling strategy in clouds,” Journal
ment option. Furthermore, we developed a novel dynamic of Parallel and Distributed Computing, vol. 75, no. 0, pp. 141 – 151,
2015.
rescheduling method and added it to handle the change [10] Z. Wen and P. Watson, “Dynamic exception handling for parti-
of cloud resources availability in our framework. This will tioned workflow on federated clouds,” in Cloud Computing Technol-
support execution resuming when clouds fail and save the ogy and Science (CloudCom), 2013 IEEE 5th International Conference
cost when new clouds become available. Additionally, we on, vol. 1, Dec 2013, pp. 198–205.
[11] R. Calheiros and R. Buyya, “Meeting deadlines of scientific work-
designed and implemented two algorithms for static de- flows in public clouds with tasks replication,” Parallel and Dis-
ployment planning, i.e., NCF and adaGA. These algorithms tributed Systems, IEEE Transactions on, vol. 25, no. 7, pp. 1787–1796,
have been applied to different types of workflows, then their July 2014.
[12] H. Hiden, S. Woodman, P. Watson, and J. Cala, “Developing cloud
performance was discussed and analysed. applications using the e-science central platform,” Royal Society
We evaluated the performance of our developed frame- of London. Philosophical Transactions A. Mathematical, Physical and
work by conducting a series of experiments on various Engineering Sciences, vol. 371, p. 20120085, 2013.
types of real scientific workflows. The experiments have [13] E. Deelman, D. Gannon, M. Shields, and I. Taylor, “Workflows
and e-science: An overview of workflow system features and
been performed using a simulation environment as well as capabilities,” Future Generation Computer Systems, vol. 25, no. 5,
real workflow management system. The results show that pp. 528–540, 2009.
our framework is suitable for deploying universal scientific [14] “Jclouds, howpublished = https://jclouds.apache.org, note = Ac-
workflows over federated clouds. cessed: 2015-09-24.”
[15] M. R. Garey and D. S. Johnson, Computers and Intractability; A Guide
As future work, we will develop a matrix to measure the to the Theory of NP-Completeness. New York, NY, USA: W. H.
security level of cloud datacenter. The existing studies only Freeman & Co., 1990.

[16] P. Watson, “A multi-level security model for partitioning work- [38] X. Fei and S. Lu, “A dataflow-based scientific workflow composi-
flows over federated clouds,” in Cloud Computing Technology and tion framework,” Services Computing, IEEE Transactions on, vol. 5,
Science (CloudCom), 2011 IEEE Third International Conference on, no. 1, pp. 45–58, Jan 2012.
2011, pp. 180–188.
[17] D. E. Bell and L. J. LaPadula, “Secure Computer Systems: Mathe- Zhenyu Wen received the B.E. degree in com-
matical Foundations,” MITRE Corporation, Tech. Rep., Mar. 1973. puter science and technology from Zhejiang
[18] C.E.Landwehr, “Formal models for computer security,” ACM Com- Gongshang University, Zhejiang, China, in 2009,
puting Surveys, vol. 13, 1981. and the M.S and Ph.D. degree in computer
[19] G. Reinelt, The Traveling Salesman: Computational Solutions for TSP science from Newcastle University, Newcastle
Applications. Berlin, Heidelberg: Springer-Verlag, 1994. Upon Tyne, U.K., in 2011 and 2015. He is
[20] Z. Wen, J. Cała, and P. Watson, “A scalable method for partitioning currently a PostDoctoral Researcher with the
workflows with security requirements over federated clouds,” in School of Informatics, the University of Edin-
Cloud Computing Technology and Science (CloudCom), 2014 IEEE 6th burgh, Edinburgh, U.K. He has authored a num-
International Conference on, Dec 2014. ber of research papers in the field of cloud com-
[21] M. Mitchell, An Introduction to Genetic Algorithms. Cambridge, puting. His current research interests include
MA, USA: MIT Press, 1998. Multi-objects optimisation, Crowdsources, Artificial Intelligent and Cloud
[22] D. Gupta and S. Ghafir, “An overview of methods maintaining computing.
diversity in genetic algorithms,” International Journal of Emerging
Technology and Advanced Engineering, vol. 2, no. 5, pp. 56–60, 2012. Rawaa Qasha is a 3rd year PhD student at the
[23] D. Bhandari, C. Murthy, and S. K. Pal, “Genetic algorithm with school of Computing Science, Newcastle Univer-
elitist model and its convergence,” International Journal of Pattern sity, UK. I received the master degree from Com-
Recognition and Artificial Intelligence, vol. 10, no. 06, pp. 731–747, puter Sciences department, University of Mosul,
1996. in 2000. Prior to start PhD career I was a lecturer
[24] R. Poli and W. B. Langdon, “Genetic programming with one-point (assistant professor) in computer science at Uni-
crossover and point mutation,” in Soft Computing in Engineering versity of Mosul, Iraq. My research interests con-
Design and Manufacturing. Springer-Verlag, 1997, pp. 180–189. centrate on Cloud computing, distributed sys-
[25] D. E. Golberg, “Genetic algorithms in search, optimization, and tem, workflow deployment, E-Science system.
machine learning,” Addion wesley, vol. 1989, p. 102, 1989.
[26] W. Sun, “Population size modeling for ga in time-critical task
scheduling,” International Journal of Foundations of Computer Science, Zequn Li received the B.E degree in Computer
vol. 22, no. 03, pp. 603–620, 2011. Science from Shandong University of Finance
[27] F. Jrad, J. Tao, I. Brandic, and A. Streit, “{SLA} enactment for large- and Economics in Shandong, China and M.S de-
scale healthcare workflows on multi-cloud,” Future Generation gree in Advanced Computer Science from New-
Computer Systems, vol. 4344, no. 0, pp. 135 – 148, 2015. castle University in Newcastle upon Tyne, United
[28] G. Juve, A. Chervenak, E. Deelman, S. Bharathi, G. Mehta, and Kingdom. Currently he is a PhD student in the
K. Vahi, “Characterizing and profiling scientific workflows,” Fu- school of Mathematics and Information Science,
ture Generation Computer Systems, vol. 29, no. 3, pp. 682 – 692, Northumbria University, United Kingdom. His re-
2013, special Section: Recent Developments in High Performance search interests include Machine Learning and
Computing and Security. Distributed Systems.
[29] J. Cała, Y. X. Xu, E. A. Wijaya, and P. Missier, “From scripted HPC-
based NGS pipelines to workflows on the cloud,” in Procs. C4Bio
workshop, co-located with the 2014 CCGrid conference. Chicago, IL: Rajiv Ranjan Dr. Rajiv Ranjan is a Associate
IEEE, 2014. Professor (Reader) in Computing Science at
[30] D. Williams, H. Jamjoom, and H. Weatherspoon, “Plug into the Newcastle University, United Kingdom. At New-
supercloud,” Internet Computing, IEEE, vol. 17, no. 2, pp. 28–34, castle University he is working on projects re-
March 2013. lated to emerging areas in parallel and dis-
[31] M. Mao and M. Humphrey, “Auto-scaling to minimize cost and tributed systems (Cloud Computing, Internet of
meet application deadlines in cloud workflows,” in High Per- Things, and Big Data). Previously, he was Julius
formance Computing, Networking, Storage and Analysis (SC), 2011 Fellow (2013-2015), Senior Research Scientist
International Conference for, Nov 2011, pp. 1–12. (equivalent to Senior Lecturer in Australian/UK
[32] M. Malawski, G. Juve, E. Deelman, and J. Nabrzyski, “Cost- and University Grading System) and Project Leader
deadline-constrained provisioning for scientific workflow ensem- in the Digital Productivity and Services Flag-
bles in iaas clouds,” in Proceedings of the International Conference on ship of Commonwealth Scientific and Industrial Research Organization
High Performance Computing, Networking, Storage and Analysis, ser. (CSIRO Australian Governments Premier Research Agency). Prior to
SC ’12. Los Alamitos, CA, USA: IEEE Computer Society Press, that he was a Senior Research Associate (Lecturer level B) in the
2012, pp. 22:1–22:11. School of Computer Science and Engineering, University of New South
[33] H. Kllapi, E. Sitaridi, M. M. Tsangaris, and Y. Ioannidis, “Sched- Wales (UNSW). Dr. Ranjan has a PhD (2009) in Computer Science
ule optimization for data processing flows on the cloud,” in and Software Engineering from the University of Melbourne. Dr. Ranjan
Proceedings of the 2011 ACM SIGMOD International Conference on is broadly interested in the emerging areas of distributed systems.
Management of Data, ser. SIGMOD ’11. New York, NY, USA: ACM, The main goal of his current research is to advance the fundamental
2011, pp. 289–300. understanding and state of the art of provisioning and delivery of appli-
[34] J. Mace, A. van Moorsel, and P. Watson, “The case for dynamic cation services (web, big data analytics, content delivery networks, and
security solutions in public cloud workflow deployments,” in De- scientific workflows) in large, heterogeneous, uncertain, and emerging
pendable Systems and Networks Workshops (DSN-W), 2011 IEEE/IFIP distributed systems.
41st International Conference on, June 2011, pp. 111–116.
[35] T. Xiaoyong, K. Li, Z. Zeng, and B. Veeravalli, “A novel security- Paul Watson is Professor of Computer Sci-
driven scheduling algorithm for precedence-constrained tasks in ence, Director of the Informatics Research Insti-
heterogeneous distributed systems,” IEEE Trans. Comput., vol. 60, tute, and Director of the North East Regional e-
no. 7, pp. 1017–1029, Jul. 2011. Science Centre. He also directs the UKRC Digi-
[36] M. HoseinyFarahabady, Y. C. Lee, and A. Zomaya, “Pareto- tal Economy Hub on Inclusion through the Digital
optimal cloud bursting,” Parallel and Distributed Systems, IEEE Economy. Prior to moving to Newcastle Univer-
Transactions on, vol. 25, no. 10, pp. 2670–2682, Oct 2014. sity, Paul worked at ICL as a system designer
[37] H. Fard, R. Prodan, and T. Fahringer, “A truthful dynamic of the Goldrush MegaServer parallel database
workflow scheduling mechanism for commercial multicloud en- server. He previously gained a Ph.D. and be-
vironments,” Parallel and Distributed Systems, IEEE Transactions on, came a Lecturer at Manchester University. His
vol. 24, no. 6, pp. 1203–1212, June 2013. research interests are in scalable information
management including data-intensive e-Science, dynamic service de-
ployment and e-Science applications. In total, Paul Watson has over
forty refereed publications, and three patents.
2168-7161 (c) 2016 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission. See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI 10.1109/TCC.2016.2603477, IEEE
Transactions on Cloud Computing
JOURNAL OF LATEX CLASS FILES, VOL. ??, NO. ??, ??? 15

Alexander Romanovsky is a Professor with

the School of Computing Science (Newcastle
University, UK) and the leader of the School’s
Secure and Resilient Systems Group. He co-
ordinates the EPSRC platform grant on Trust-
worthy Ambient Systems and the RSSB Safe-
Cap+ grant on railway integrated optimum ca-
pacity, safety and energy strategies. He is a co-
investigator of the EPSRC PRiME programme
grant on Power-efficient, Reliable, Many-core
Embedded systems. Before this, he coordinated
the major EU FP7 DEPLOY IP that developed the Rodin tooling environ-
ment for formal stepwise design of complex dependable systems using
Event-B. Rodin is now widely used by companies in Europe, China,
Japan, USA, Canada and Brazil. Prof Romanovsky’s main research
areas are system dependability, fault tolerance, safety, modelling and
verification.

Grade 7 English
No ratings yet
Grade 7 English
12 pages
Maths Class Ix Chapter 04 05 and 06 Practice Paper 02
100% (1)
Maths Class Ix Chapter 04 05 and 06 Practice Paper 02
4 pages
Jamgon Mipham Treasury of Blessings Practice
100% (2)
Jamgon Mipham Treasury of Blessings Practice
10 pages
A Considerable Speck ICSE Class 10 English Answers, Notes
No ratings yet
A Considerable Speck ICSE Class 10 English Answers, Notes
3 pages
1100DW Manual Guide V2.0
No ratings yet
1100DW Manual Guide V2.0
24 pages
Systems Engineering Management Plan
No ratings yet
Systems Engineering Management Plan
27 pages
Knowing The Different Types of Film Genre
No ratings yet
Knowing The Different Types of Film Genre
13 pages
HIS-Print culture-MCQs-QUES
No ratings yet
HIS-Print culture-MCQs-QUES
8 pages
Schopenhauer, Arthur Singh, R. Raj Death, Contemplation and Schopenhauer PDF
100% (1)
Schopenhauer, Arthur Singh, R. Raj Death, Contemplation and Schopenhauer PDF
141 pages
Lesson 1.a - 1.c - Integration Concepts & Formulas
No ratings yet
Lesson 1.a - 1.c - Integration Concepts & Formulas
5 pages
Isomorphisms and Allomorphisms in The Morphemic Structure of English and Ukrainian Words
No ratings yet
Isomorphisms and Allomorphisms in The Morphemic Structure of English and Ukrainian Words
13 pages
PurpComm Unit 1 Lesson 3
No ratings yet
PurpComm Unit 1 Lesson 3
34 pages
Int 21
No ratings yet
Int 21
7 pages
8086 Hardware Specification
100% (1)
8086 Hardware Specification
84 pages
Basic Principles of Turbomachines
No ratings yet
Basic Principles of Turbomachines
41 pages
Elements of Potery
No ratings yet
Elements of Potery
10 pages
Evolve 1 Unit 1 PPT Lesson 3
No ratings yet
Evolve 1 Unit 1 PPT Lesson 3
9 pages
PCB-Gcode-Wizard Version 100 Help File Manual en Ingles
No ratings yet
PCB-Gcode-Wizard Version 100 Help File Manual en Ingles
9 pages
Hallowed Be Your Name
No ratings yet
Hallowed Be Your Name
23 pages
Class 2 Eng QB 2025-26
No ratings yet
Class 2 Eng QB 2025-26
14 pages
Lost/Found in Translation: Qurratulain Hyder As Self-Translator
No ratings yet
Lost/Found in Translation: Qurratulain Hyder As Self-Translator
16 pages
Be Your Best Self
No ratings yet
Be Your Best Self
7 pages
Prais Winsten Regression
No ratings yet
Prais Winsten Regression
33 pages
United Independent Bengal
No ratings yet
United Independent Bengal
3 pages
Difficulties in Listening of The First Year Students at Tay Do University in Vietnam
No ratings yet
Difficulties in Listening of The First Year Students at Tay Do University in Vietnam
9 pages
Q & A Exam (Adv V11) Q & A Exam (Adv V11) : Review Your Answers
No ratings yet
Q & A Exam (Adv V11) Q & A Exam (Adv V11) : Review Your Answers
7 pages
Aaquib Resume
No ratings yet
Aaquib Resume
2 pages
Project Report PDF
No ratings yet
Project Report PDF
5 pages
CHEM201 Slides 7
No ratings yet
CHEM201 Slides 7
4 pages
Rubrik Penilaian Pembentangan
No ratings yet
Rubrik Penilaian Pembentangan
3 pages
Principles: Life and Work
From Everand
Principles: Life and Work
Ray Dalio
4/5 (648)
The Glass Castle: A Memoir
From Everand
The Glass Castle: A Memoir
Jeannette Walls
4.5/5 (1856)
The Emperor of All Maladies: A Biography of Cancer
From Everand
The Emperor of All Maladies: A Biography of Cancer
Siddhartha Mukherjee
4.5/5 (298)
The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are
From Everand
The Gifts of Imperfection: Let Go of Who You Think You're Supposed to Be and Embrace Who You Are
Brené Brown
4/5 (1175)
Steve Jobs
From Everand
Steve Jobs
Walter Isaacson
4.5/5 (1139)
The Perks of Being a Wallflower
From Everand
The Perks of Being a Wallflower
Stephen Chbosky
4.5/5 (4103)
Shoe Dog: A Memoir by the Creator of Nike
From Everand
Shoe Dog: A Memoir by the Creator of Nike
Phil Knight
4.5/5 (629)
Angela's Ashes: A Memoir
From Everand
Angela's Ashes: A Memoir
Frank McCourt
4.5/5 (943)
Rise of ISIS: A Threat We Can't Ignore
From Everand
Rise of ISIS: A Threat We Can't Ignore
Jay Sekulow
3.5/5 (144)
The Outsider: A Novel
From Everand
The Outsider: A Novel
Stephen King
4/5 (2886)
The Yellow House: A Memoir (2019 National Book Award Winner)
From Everand
The Yellow House: A Memoir (2019 National Book Award Winner)
Sarah M. Broom
4/5 (100)
The Light Between Oceans: A Novel
From Everand
The Light Between Oceans: A Novel
M.L. Stedman
4.5/5 (815)
Team of Rivals: The Political Genius of Abraham Lincoln
From Everand
Team of Rivals: The Political Genius of Abraham Lincoln
Doris Kearns Goodwin
4.5/5 (244)
The World Is Flat 3.0: A Brief History of the Twenty-first Century
From Everand
The World Is Flat 3.0: A Brief History of the Twenty-first Century
Thomas L. Friedman
3.5/5 (2289)
Fear: Trump in the White House
From Everand
Fear: Trump in the White House
Bob Woodward
3.5/5 (836)
Sing, Unburied, Sing: A Novel
From Everand
Sing, Unburied, Sing: A Novel
Jesmyn Ward
4/5 (1267)
The Unwinding: An Inner History of the New America
From Everand
The Unwinding: An Inner History of the New America
George Packer
4/5 (45)
Her Body and Other Parties: Stories
From Everand
Her Body and Other Parties: Stories
Carmen Maria Machado
4/5 (903)
Little Women
From Everand
Little Women
Louisa May Alcott
4.5/5 (2369)
A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story
From Everand
A Heartbreaking Work Of Staggering Genius: A Memoir Based on a True Story
Dave Eggers
3.5/5 (233)
John Adams
From Everand
John Adams
David McCullough
4.5/5 (2546)
Manhattan Beach: A Novel
From Everand
Manhattan Beach: A Novel
Jennifer Egan
3.5/5 (919)

Dynamically Partitioning Workflow Over Federated Clouds For Optimising The Monetary Cost and Handling Run-Time Failures

Uploaded by

Dynamically Partitioning Workflow Over Federated Clouds For Optimising The Monetary Cost and Handling Run-Time Failures

Uploaded by

This article has been accepted for publication in a future issue of this journal, but has not been

Dynamically Partitioning Workflow over

S CIENTIFIC workflows have become an increasingly pop-

In addition, Λ0 is the set of possible deployments of the

Dcost(λ0 ) = COST (λ0 ) + Icost(Selected) (6)

Where λ0 ∈ Λ0 represents one of the possible deploy-

ALGORITHM 1: Elitist Prevention

4.4 Adaptive Rescheduling Algorithm

ALGORITHM 5: Dynamic rescheduling ALGORITHM 6: Redeploy Services

the assigned tasks before their terminating time. (2) SBR:

To simulate Availability of new clouds, the workflow has to

NCF adaGA HUGA

Alexander Romanovsky is a Professor with

You might also like