Model-based Adaptation of Cloud Computing Applications

Christian Inzinger, Benjamin Satzger, Philipp Leitner, Waldemar Hummer and Schahram Dustdar

Distributed Systems Group, Vienna University of Technology,

Argentinierstrasse 8/184, A-1040 Vienna, Austria

Keywords:

Cloud Computing, Model-based Adaptation.

Abstract:

In this paper we propose a provider-managed, model-based adaptation approach for cloud computing applica-

tions, allowing customers to easily specify application behavior goals or adaptation rules. Delegating control

over corrective actions to the cloud provider will pose advantages for both, customers and providers. Cus-

tomers are relieved of effort and expertise requirements necessary to build sophisticated adaptation solutions,

while providers can incorporate and analyze data from a multitude of customers to improve adaptation de-

cisions. The envisioned approach will enable increased application performance, as well as cost savings for

customers, whereas providers can manage their infrastructure more efﬁciently.

1 INTRODUCTION

The cloud computing paradigm has found widespread

adoption throughout the last couple of years. The

pay-per-use model of cloud computing proved to be

convenient in many respects. It allows to cope with

services of time-varying resource demand and peak

loads. Cloud computing helps to avoid high initial

investments in IT infrastructures, as resources can be

dynamically provisioned even if the demand for a ser-

vice is unknown in advance (Armbrust et al., 2010).

However, in order to beneﬁt from the resource

elasticity provided by cloud computing, applications

need to be properly built. In this paper we argue

that developers should be able to create models de-

scribing their application’s adaptation capabilities to-

gether with adaptation goals deﬁning their prefer-

ences. Cloud computing providers use this informa-

tion to control and adapt the application according

to the customers’ objectives. Customers have only

a limited view on the execution infrastructure, while

cloud computing providers, when given access to nec-

essary information from the application, have a com-

plete view and can incorporate low-level infrastruc-

ture details to make adaptation more efﬁcient and ef-

fective.

Models are already used for improving deploy-

ment of cloud computing applications. An example

is Amazon’s service called CloudFormation

, provid-

http://aws.amazon.com/cloudformation

ing a domain speciﬁc language (DSL) for deﬁning

the infrastructure required by an application, which

can be provisioned in a single step. Research has

brought forth sophisticated approaches for adapt-

ing application topologies and resource allocations,

e.g., for latency requirements (Chang et al., 2012),

power and cost considerations (Jung et al., 2010),

or deadline-driven workﬂow scheduling (Kim et al.,

2010), among other purposes. In practice, however,

cloud providers currently do not provide mechanisms

for further managing an application at runtime. We

propose to use models allowing application devel-

opers to create “management hooks” for the cloud

provider. This has the advantages that all applications

can beneﬁt from a sophisticated control mechanisms

offered by the provider. Application developers use

models to state the objectives for application control,

relieving them from implementing complex adapta-

tion schemes. Cloud providers have a deeper under-

standing of the application infrastructure, which they

can use to improve application adaptation, but also to

optimize resource usage. Also, providers can leverage

the experiences and data from similar applications to

improve control and adaptation over time.

In the next section we point out how models are

used in cloud computing today. Then, in Section 3

we present our model-based adaptation approach. We

wrap up in Section 4 and provide an outlook for future

research and remaining challenges.

351

Inzinger C., Satzger B., Leitner P., Hummer W. and Dustdar S..

Model-based Adaptation of Cloud Computing Applications.

DOI: 10.5220/0004381803510355

In Proceedings of the 1st International Conference on Model-Driven Engineering and Software Development (MODA-2013), pages 351-355

ISBN: 978-989-8565-42-6

 2013 SCITEPRESS (Science and Technology Publications, Lda.)

2 MODELS IN CLOUD

COMPUTING

The main responsibilities of cloud application lifecy-

cle management are infrastructure provisioning, ap-

plication deployment, and ﬁnally control and adapta-

tion at runtime.

Infrastructure provisioning involves setting up vir-

tual machines, installing and conﬁguring required

software packages, and initializing cloud services,

such as load balancers, database instances, persis-

tent storage, or caches. Manual infrastructure pro-

visioning is cumbersome and error-prone, and limits

the level of reusability, e.g., to set up identical envi-

ronments for development and testing. Furthermore,

proper change management of infrastructure is not

supported. Model-based approaches for infrastructure

provisioning that have recently emerged help to over-

come these limitations. Amazon’s CloudFormation

provides a service to create infrastructure templates

used to create a collection of related infrastructure

resources and provision them in an automated way.

These models can be easily reused and shared using

infrastructure template repositories, helping to estab-

lish and distribute best practices for different kinds

of applications. The provisioned resources, e.g., vir-

tual machines, still need to be properly conﬁgured

to include all necessary application dependencies, as

well as respective links between system components.

To enable a predictable and repeatable process for

this, deployment models are used to enable automated

server conﬁguration according to predeﬁned speciﬁ-

cations. Popular approaches for modeling software

deployment include conﬁguration management (CM)

tools like Chef

, Puppet

, or cfengine

. Resource

conﬁguration is modeled using vendor-speciﬁc DSLs,

specifying required software packages and libraries,

as well as necessary conﬁguration ﬁles and parame-

ters to support the application to be deployed and all

its components. The CM will attempt bring all man-

aged resources into the state deﬁned for its particular

role by installing packages, deploying conﬁguration

ﬁles, and (re-)starting affected services.

Application deployment and redeployment deal

with packaging and copying artifacts to according

infrastructure resources, conﬁguring them, establish-

ing links between them, and ﬁnally making sure

they are properly launched. Deployment models are

used to ensure efﬁcient and effective application roll-

outs and updates based on declarative speciﬁcations,

http://opscode.com/chef/

http://puppetlabs.com

http://cfengine.com

Cloud Provider Infrastructure

Auto Scaling Gr.

Load Balancer

Web Server

Load Balancer

Auto Scaling Gr.

App Server

Adaptation

DB Master

DB Read

Slave

Users

component interactionadaptation logic adaptation interaction

Figure 1: Traditional Cloud Application Architecture.

e.g. (Inzinger et al., 2012). Again, CM tools, as men-

tioned above, can be used to model application com-

ponents, their dependencies, and required parameter

settings.

While deployment models enable the predictable,

testable and repeatable provisioning of runtime in-

frastructure, their responsibility stops at deployment

time. The models are not designed to consider ap-

plication speciﬁc runtime aspects or application be-

havior goals that should be reached. Cloud providers

offer means to monitor basic infrastructure or ap-

plication metrics, but adaptation is mostly limited

to starting or stopping instances to cope with vary-

ing load patterns. Complex applications, however,

have a variety of possibilities to cope with changes to

the environment without necessarily resorting to scal-

ing out available resources. Certain non-critical pro-

cess steps could, for instance, be deferred to a later

time if current load is too high, or service quality

could be adapted according to environmental circum-

stances. Today’s cloud providers, however, employ

a black box model for applications deployed on their

infrastructure, leaving customers responsible for re-

alizing their own adaptation infrastructure to imple-

ment application-level reactions to changes in the en-

vironment. Future cloud service providers will offer

means for integrating advanced application adapta-

tion into their offerings using a model-based adapta-

tion approach, as discussed in the next section.

3 A CASE FOR MODEL-BASED

ADAPTATION

In this section, we present an approach for provider-

assisted model-based application adaptation that will

yield beneﬁts for both providers and customers.

Figure 1 shows a traditional cloud application de-

ployment for an exemplary web application. Incom-

ing requests are distributed across several front end

web servers using the provider’s load balancer ser-

vice. The web servers access the back end infrastruc-

MODELSWARD2013-InternationalConferenceonModel-DrivenEngineeringandSoftwareDevelopment

352

Auto Scaling Group

Cloud Provider Infrastructure

Auto Scaling Group

Load Balancer

Web Server Load Balancer App Server

DB Master

DB Read

Slave

Application Objectives

Users

component interactionadaptation modeladaptation point

Figure 2: Cloud Application Architecture using Provider-Managed Adaptation.

ture using a custom middle tier load balancer, evenly

distributing load on the available back end applica-

tion servers. The application servers execute business

logic and access the database cluster. Web and appli-

cation server instances can be added or removed ac-

cording to their current and predicted future load. The

custom adaptation component controls service qual-

ity by performing corrective actions on all application

services according to deﬁned behavior goals, realized

as autonomic control loop (Huebscher and McCann,

2008). It will, for instance, incrementally raise the

fragment cache expiry time out up to a certain thresh-

old on the web servers when an increase in request

trafﬁc is detected. Similarly, a recommendation mod-

ule running on the app server instances will set to only

return cached user group recommendations, or a list

of the most purchased products, instead of personal-

ized recommendations, in response to high order vol-

umes. This will lower stress on the database, as well

as the application servers and starting new instances is

not immediately required. Analogously, when appli-

cation load is low, the web servers will always serve

fresh content and the recommendation module will al-

ways deliver personalized recommendations. Since

these adaptation actions inﬂuence service quality and

thus user experience, the customer will employ a util-

ity function within the adaptation component to de-

termine, when to change service quality and when

to scale. Currently, cloud providers only offer sup-

port for automatic scaling of application components

based on infrastructure and application metrics, but,

as deployed applications are considered black boxes,

do not provide means to alter component runtime be-

havior to react to changes in the environment. As in-

dicated in the ﬁgure, customers need to deploy their

own adaptation infrastructure, often scattered across

deployed components.

We propose an approach for model-based cloud

application adaptation. Customers specify emit-

ted metrics, available adaptation points and desired

objectives using provided models, which could be

based on previous research such as (Zhang and

Cheng, 2006), in order to make applications ready

for provider-driven adaptation and control. This gives

customers the ability to declaratively specify desired

application behavior, while the provider has neces-

sary information to take over control. Figure 2 shows

the cloud application discussed above, deployed using

the envisioned model-based provider-managed adap-

tation approach. A dedicated custom adaptation com-

ponent is not needed anymore, as customers describe

desired application behavior objectives in the adap-

tation model and delegate its execution to the cloud

provider. This results in signiﬁcant cost savings for

the customer, as creating sophisticated and compre-

hensive adaptation solutions requires considerable ef-

fort and expertise. The business logic components

do not contain any adaptation logic, but only pro-

vide designated adaptation points to allow for ex-

ternal management. Available adaptation points are

referenced in the adaptation model for use in cor-

rective actions. As a result, component design is

simpliﬁed through the clear separation of adaptation

capabilities and adaptation logic. As desired appli-

cation behavior is declaratively speciﬁed in the ap-

plication adaptation model, it can easily be reused

or modiﬁed to accommodate changes without alter-

ing components. In the presented exemplary appli-

cation, the customer speciﬁes the following adapta-

tion points: SetFragmentCacheTimeout to modify

the fragment cache timeout in the web server com-

ponent, and SetRecommendationStrategy to switch

between personalized recommendations, cached cus-

tomer group recommendations, or a list of the most

sold products overall in the application server. Addi-

tionally, the ScalePool adaptation point supplied by

Model-basedAdaptationofCloudComputingApplications

353

the cloud provider can be used to start and stop ap-

plication instances. Monitoring metrics data required

to take adaptation decisions are modeled with the ap-

plication objectives, or a push-based approach simi-

lar to current provider-assisted scaling solutions such

as the previously Amazon CloudWatch is used. The

model allows for speciﬁcation of conventional event

condition action (ECA) rules to govern application

behavior, so existing applications can easily migrate

their current adaptation strategy. Alternatively, the

customer can deﬁne a simple utility function to re-

ﬂect application objectives according to desired char-

acteristics, such as ‘response time should not exceed

500ms’, ‘service quality should be as high as possi-

ble’, and ‘infrastructure cost should be as low as pos-

sible’. This is possible since adaptation points contain

indications on how they will affect application behav-

ior. Adaptation point SetFragmentCacheTimeout,

for instance, will indicate that higher parameter values

should decrease response time, but will also decrease

service quality. These indications will initially be sup-

plied by customers, but the cloud provider will mon-

itor the effects of performed corrective actions and

adjust adaptation point information during the run-

time of the application. The provider-managed adap-

tation service uses the customer-supplied objectives

and infers optimized actionable adaptation strategies.

A cloud provider can not only use monitoring data

from the application to be controlled, but also con-

sider historical data from different but similar cus-

tomers, leveraging time series analysis to better an-

ticipate load patterns or even adaptation action effects

for common components such as databases.

Allowing for provider-assisted application adap-

tation at runtime has a number of distinct advantages.

Outsourcing some control over adaptation decisions

will beneﬁt customers. Providers can offer optimized

adaptation decisions based on data from similar cus-

tomers. Furthermore, providers are able to improve

resource provisioning strategies and offer better ser-

vice to customers.

By handing off control over the execution of

adaptation actions, customers can beneﬁt from the

provider’s expertise in reliability and resilience en-

gineering. They are relieved of implementing their

own complex adaptation mechanisms and can use an

advanced, thoroughly tested solution offered by the

cloud provider without investing large amounts of

capital or personnel. Leveraging economies of scale,

the cloud provider in turn can offer their adaptation

solution to customers with competitive pricing strate-

gies while retaining appropriate return on investment.

Using data from a multitude of customers, the

cloud provider can make better adaptation decisions

for customers, enabling higher levels of service, as

well as cost savings for customers and the provider.

By analyzing time series data of a large number of

customers, the provider can anticipate, for instance,

future load patterns much more accurately than a sin-

gle customer could with their data alone. This allows

for better adaptation decisions, as the provider can es-

tablish categories of customers, enabling more pre-

cise load predictions based on signals like trafﬁc pat-

terns and geographical distribution. Using this data,

the provider can offer targeted adaptation decisions

or recommendations for geographic placement of in-

stances, or even migration of instances to more suit-

able locations.

Furthermore, the additional information available

to the provider will enable highly optimized resource

provisioning, keeping over-provisioning at even lower

levels than with traditional elastic applications that

can only analyze trafﬁc patterns from their own past.

Customers naturally beneﬁt from lower resource us-

age by paying less, while the provider now has ac-

cess to previously unused but, due to inefﬁcient provi-

sioning, inaccessible resources. Furthermore, the sup-

plied adaptation goals can be used to react in different

ways to varying environmental conditions. Depend-

ing on the utility, it might be more reasonable for an

application to defer certain processing steps to a later

time, thus reducing load on application instances, be-

fore scaling out by adding new machines. Similarly,

when request trafﬁc declines it might be beneﬁcial

to process queued tasks on the now underused in-

stances – at least until their current billing interval is

over – before terminating them to reduce infrastruc-

ture costs. The customer-provided utility functions

guide the adaptation decisions considering applica-

tion performance, service quality, as well as infras-

tructure costs. While the presented approach is ide-

ally deployed by the cloud provider, a transitional im-

plementation could also be realized as part of a cloud

abstraction layer like the meta cloud (Satzger et al.,

2013), allowing the use of a managed adaptation so-

lution without explicit cloud provider support.

4 CONCLUSIONS

In this paper we presented a novel approach for

provider-managed, model-based adaptation of cloud

computing applications. Customers are not required

to implement their own adaptation solution, but can

use a simple model to specify desired application

behavior. The presented approach has the poten-

tial for signiﬁcant cost savings and increased appli-

cation quality for customers by utilizing a sophisti-

MODELSWARD2013-InternationalConferenceonModel-DrivenEngineeringandSoftwareDevelopment

354

cated adaptation infrastructure managed by the cloud

provider that can offer better adaptation decisions by

considering data from multiple different, but similar,

customers. Furthermore, cloud providers will be able

to manage their infrastructure more efﬁciently due

to reduced resource over-provisioning resulting from

improved adaptation decisions.

Deferring control over adaptation decisions to a

cloud provider poses several challenges that need to

be addressed. Our approach requires that customers

trust providers with potentially conﬁdential informa-

tion about their applications’ inner structure, such as

adaptation strategies and request trafﬁc patterns. The

widespread adoption of cloud computing has shown

that customers already trust reputable providers with

hosting their applications, so we argue that the addi-

tion of adaptation strategies will not be problematic,

provided that service contracts clearly state how pro-

vided data will be used. Customers that do not want

their data used to improve adaptation decisions for

others could still beneﬁt of the provider’s adaptation

infrastructure, saving them implementation effort, but

will in turn not be able to receive adaptation decisions

optimized using data from others. Furthermore, ser-

vice level agreements need to explicitly state the new

DSLs and tools needed, as well as responsibilities of

both customers and providers when using managed

adaptation infrastructure.

ACKNOWLEDGEMENTS

The research leading to these results has received

funding from the European Commission’s Seventh

Framework Program [FP7/2007-2013] under grant

agreement 257483 (Indenica), as well as from the

Austrian Science Fund (FWF) under grant P23313-

N23 (Audit 4 SOAs).

REFERENCES

Armbrust, M., Stoica, I., Zaharia, M., Fox, A., Grifﬁth, R.,

Joseph, A. D., Katz, R., Konwinski, A., Lee, G., Pat-

terson, D., and Rabkin, A. (2010). A View of Cloud

Computing. Communications of the ACM, 53(4):50–

58.

Chang, F., Viswanathan, R., and Wood, T. L. (2012). Place-

ment in Clouds for Application-Level Latency Re-

quirements. In 5th IEEE International Conference on

Cloud Computing (CLOUD), pages 327–335.

Huebscher, M. C. and McCann, J. A. (2008). A survey of

autonomic computing—degrees, models, and applica-

tions. ACM Computing Surveys, 40(3):7:1–7:28.

Inzinger, C., Satzger, B., Hummer, W., and Dustdar, S.

(2012). Speciﬁcation and deployment of distributed

monitoring and adaptation infrastructures. In Work-

shop on Performance Assessment and Auditing in Ser-

vice Computing (PAASC) held in conjunction with IC-

SOC.

Jung, G., Hiltunen, M., Joshi, K., Schlichting, R., and Pu, C.

(2010). Mistral: Dynamically managing power, per-

formance, and adaptation cost in cloud infrastructures.

In 30th IEEE International Conference on Distributed

Computing Systems (ICDCS), pages 62–73.

Kim, H., el Khamra, Y., Jha, S., and Parashar, M. (2010).

Exploring application and infrastructure adaptation

on hybrid grid-cloud infrastructure. In 19th ACM

International Symposium on High Performance Dis-

tributed Computing (HPDC), pages 402–412.

Satzger, B., Hummer, W., Inzinger, C., and Dustdar, S.

(2013). Winds of Change: From Vendor Lock-In to

the Meta Cloud. Internet Computing, 17(1).

Zhang, J. and Cheng, B. H. C. (2006). Model-based

development of dynamically adaptive software. In

28th International Conference on Software engineer-

ing (ICSE), pages 371–380.

Model-basedAdaptationofCloudComputingApplications

355