TOWARDS PERFORMANCE PREDICTION FOR CLOUD

COMPUTING ENVIRONMENTS BASED ON

GOAL-ORIENTED MEASUREMENTS

Michael Hauck

FZI Research Center for Information Technology, Karlsruhe, Germany

Jens Happe

SAP Research, Karlsruhe, Germany

Ralf H. Reussner

Karlsruhe Institute of Technology (KIT), Karlsruhe, Germany

Keywords:

Performance prediction, Measurements, Cloud computing, Virtualization, Modelling.

Abstract:

Scalability and performance are critical quality attributes of applications developed for the cloud. Many of

these applications have to support hundreds or thousands of concurrent users with strongly ﬂuctuating work-

loads. Existing approaches for software performance evaluation do not address the new challenges that arise

for applications executed in cloud computing environments. The effects of virtualization on response times,

throughput, and resource utilisation as well as the massive number of resources available require new platform

and resource models for software performance evaluation. Modelling cloud environments using established

approaches for software performance prediction is a cumbersome task that requires a detailed understanding

of virtualization techniques and their effect on software performance. Additional complexity comes from the

fact that cloud environments may combine multiple virtualization platforms which differ in implementation

and performance properties.

In this position paper, we propose an approach to infer performance models of cloud computing environments

automatically through goal-oriented measurements. The resulting performance models can be directly com-

bined with established model-driven performance prediction approaches. We outline the research challenges

that have to be addressed in order to employ the approach for design-time performance predictions of software

systems running in cloud computing environments.

1 INTRODUCTION

Many applications developed for the cloud have high

requirements with respect to performance and scal-

ability in order to support hundreds or thousands of

users. Response time and throughput of such appli-

cations are critical and thus have to be considered in

early stages of the software life-cycle. Software ar-

chitects must take into account the speciﬁc proper-

ties of cloud environments. Most fundamental is a

detailed understanding of different virtualization plat-

forms and their effect on software performance. Fur-

thermore, virtualization platforms provide means to

scale up (increase the size of a virtual machine by

assigning it more memory or cpus) and scale out

(adding additional virtual machines to an application).

Both must be supported by the application. Perfor-

mance standards in practice are high. An application

is considered scalable only if its throughput increases

linearly with the resources assigned to it.

Software performance engineering (SPE) and

model-driven performance engineering (MDPE) en-

able the prediction of software performance at an

early stage of the software development process.

Software architects provide architectural models with

performance annotations of their application, which

are then transformed into performance analysis mod-

els, such as queueing networks, stochastic Petri nets,

or simulation models (see overview in (Balsamo et al.,

2004; Koziolek, 2010)). The results of the perfor-

mance analysis enable software architects to reason

on the application’s performance, for example to eval-

616

Hauck M., Happe J. and Reussner R..

TOWARDS PERFORMANCE PREDICTION FOR CLOUD COMPUTING ENVIRONMENTS BASED ON GOAL-ORIENTED MEASUREMENTS.

DOI: 10.5220/0003387406160622

In Proceedings of the 1st International Conference on Cloud Computing and Services Science (CLOSER-2011), pages 616-622

ISBN: 978-989-8425-52-2

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

uate different design alternatives, to plan resource ca-

pacities or to estimate the performance of a software

system in a cloud.

For accurate performance predictions, the

performance-relevant properties of cloud environ-

ments have to be reﬂected by the performance

models. However, current approaches provide only

limited support for performance properties of cloud

computing environments and, more speciﬁcally,

virtualization platforms. Therefore, the integration

of such properties is a cumbersome, manual task,

which requires detailed knowledge about virtu-

alization techniques and their effect on software

performance. In many cases, performance-relevant

properties of virtualization platforms have to be

retrieved by inspecting system documentation. Such

documentation may not be available in all cases.

Furthermore, the effect of speciﬁc implementation

details on performance is often unknown. However,

neglecting the performance impact of virtualization

platforms can lead to inaccurate prediction results.

For example, scheduling implementations can inﬂu-

ence the response time of an application by several

orders of magnitude (Schroeder et al., 2006). If

no information about the virtualization platform is

available, measurements have to be conducted in

order to include its inﬂuence into the performance

model. Such measurements involve a high manual

effort to design and set up the measurements and

analyze the results.

Moreover, cloud computing environments can

make use of multiple virtualization techniques,

which differ in their performance-relevant proper-

ties (Adams and Agesen, 2006), (Matthews et al.,

2007). Thus, a performance model of a speciﬁc vir-

tualization platform cannot necessarily be reused in a

different setting.

To avoid the efforts of creating the performance

model manually, we propose an approach to detect

performance-relevant properties of cloud computing

environments and virtualization environments auto-

matically. This approach makes use of goal-oriented

measurements which are conducted on the target plat-

form to detect certain characteristics of performance-

relevant properties of the virtualization environment.

The environment is regarded as a “black-box”, which

means that the performance-relevant properties of the

environment are not known a priori. Instead, the prop-

erties are derived by measurements that are executed

on the platform, and fed into a performance model.

The automation of both the measurements and the

derivation of a performance model aims at easing the

burden of performance predictions of software run-

ning in the cloud.

The remainder of this position paper is structured

as follows. Section 2 discusses research questions re-

garding performance prediction of software running

in cloud computing or virtualized environments based

on measurements. The approach is presented in detail

in Section 3. Section 4 presents related work, and

Section 5 concludes the paper.

2 RESEARCH QUESTIONS

In this section, several open research questions are

presented that need to be addressed in order to apply

performance predictions for cloud computing envi-

ronments based on goal-oriented measurements. The

ﬁrst three questions deal with design-based perfor-

mance prediction for software running in virtual-

ized environments, the latter two questions deal with

goal-oriented measurements to derive performance-

relevant properties.

What are the Relevant Performance Properties of

Virtualization Platforms?

Running software in virtualized environments leads to

a complex technology stack which inﬂuences the soft-

ware performance. To enable reasonable performance

analyses, the key performance-relevant properties of

virtualization environments have to be identiﬁed.

Such properties include for example virtualization

overheads and virtualization implementation proper-

ties. Virtualization overhead has to be taken into

account, as resource demands of software may lead

to a different performance if the software is run-

ning on a virtual machine compared to its execution

in a non-virtualized environment. Virtualization im-

plementation properties play for example a crucial

role in handling contention effects of multiple vir-

tual machines running on the same physical server.

This is done by a virtualization software called hy-

pervisor, which provides sophisticated scheduling al-

gorithms that affect the performance. Even for the

same virtualization technique, different scheduler im-

plementations may lead to different performance ef-

fects (Cherkasova et al., 2007). Hypervisors also pro-

vide a more complex logic to handle I/O requests

compared to traditional OS schedulers, as virtual I/O

devices have to be mapped to physical I/O devices.

Further performance-relevant properties that have

been identiﬁed for software running in a cloud com-

puting environment include network I/O properties.

For example, some parts of a software may run locally

with a local network connection, whereas other parts

may run in the cloud where internet network connec-

tion properties have to be regarded.

In addition to virtualized platforms, cloud com-

TOWARDS PERFORMANCE PREDICTION FOR CLOUD COMPUTING ENVIRONMENTS BASED ON

GOAL-ORIENTED MEASUREMENTS

617

puting environments hide the server structure on

which the application is running on. Even more, af-

ter deploying an application in the cloud, it may be

moved dynamically to a different server with differ-

ent performance properties after deployment. In our

presented approach we want to focus on scenarios

where the application is running on the same infras-

tructure on which the black-box measurements have

been taken. However, we acknowledge the fact that

cloud computing environments may imply more com-

plex inﬂuencing factors for performance modelling,

which are subject to later enhancements of the ap-

proach.

How to Reﬂect Relevant Performance Properties

of Cloud Computing Environments/Virtualization

Platforms in Software Performance Models?

For design-time performance prediction, models of

the software have to be provided, and to some ex-

tent models of the virtualized environment. Here, the

granularity of the model mainly depends on the ap-

proach to be taken:

In a forward-engineering approach, the implemen-

tation of the application often does not exist yet, and

the models tend to be rather abstract. Here, the chal-

lenge is to ﬁnd a model granularity that does not re-

quire knowledge which not available during the de-

sign phase, but still allows to yield reasonable per-

formance prediction results. Another option would

be to use abstract models and automatically add addi-

tional information by using performance completions

or coupled transformations, as proposed in (Happe

et al., 2010).

In a reverse-engineering approach, the code of

the software can be taken into account when creat-

ing the models. This allows to create more ﬁne-

grained models, which can be partially created au-

tomatically by tool support. Here, also more ﬁne-

grained performance-relevant information can be put

into the model. Such an approach might be relevant

for a scenario in which the performance of legacy sys-

tems should be analyzed w. r. t. a deployment in a vir-

tualized environment.

Independent of the software model granularity, ex-

isting models for performance predictions have to be

enhanced in order to reﬂect virtualization and cloud

computing infrastructure properties, such as virtual-

ization hypervisor properties, different network con-

nections, virtual and physical servers, or server clus-

ters.

How to Integrate Performance-relevant

Properties of Cloud Computing Environments

into Performance Analyses?

Additionally to the enhancements of software perfor-

mance models, performance analysis tools have to be

enhanced as well in order to support performance-

relevant properties of cloud computing environments.

For performance prediction, we aim at using a con-

ﬁgurable analysis tool that can be conﬁgured depend-

ing on the detected properties (e.g. performance over-

heads or hypervisor properties). As the complete ap-

proach should run in an automated way where possi-

ble, analysis tool conﬁguration should also happen in

an automated way.

To implement performance prediction for our ap-

proach, we plan to enhance the Palladio Compo-

nent Model (PCM) (Becker et al., 2009). The PCM

supports design-based performance prediction based

on software models and provides different perfor-

mance analysis tools, such as analytical solvers and

a discrete-event simulation.

How to Design Experiments to Detect

Performance-relevant Properties in a

Technology-independent Way?

In order to apply performance prediction to differ-

ent cloud environments, the measurements of the ex-

periments have to be designed in a generic way and

must not be tailored towards a speciﬁc virtualization

technique or hypervisor scheduler implementation. It

is also an open question which measurement metrics

can be taken into account. While every system pro-

vides means to measure time spans, some virtualiza-

tion solutions or cloud computing providers might not

provide certain measurement metrics such as resource

utilizations.

In (Hauck et al., 2010), we proposed generic

experiments to detect load-balancing properties of

general-purpose operating system (GPOS) sched-

ulers. The experiments are based on response time

measurements only and yielded promising results.

Furthermore, measurements aiming at detecting

the virtualization overhead of certain resource de-

mands (CPU, memory, network I/O) have already

been conducted in prior work. We plan to automate

the evaluation of the measurements in order to auto-

matically derive performance models from the mea-

surement results which can be used for performance

prediction.

Depending on the scenario, different additional

kinds of experiments might be possible. For exam-

ple, one could think of taking measurements on mul-

tiple machines in the cloud to gain experiment results

that allow to reason about the contention effects that

occur.

CLOSER 2011 - International Conference on Cloud Computing and Services Science

618

Deploy Drivers

Perform

Measurements

Derive

Performance-

relevant

Properties

Conduct

Performance

Analysis

Integrate

Derived

Properties into

Analysis Tool

Create

Software

Performance

Model

Figure 1: Workﬂow of the automatic derivation of performance-relevant properties.

How to Conduct Accurate Measurements in

Virtualized Environments?

In virtualized environments, additional challenges

arise when taking measurements. First, the timer ac-

curacy might suffer when accessing timer information

from inside a virtual machine. Virtualization imple-

mentations provided enhanced timer accuracy in the

last years, but initial experiment results showed that

timer accuracy might suffer when measuring response

time in virtual machines especially when the system

is under load.

Another issue when taking measurements is that

additional load might disturb the measurement re-

sults. In virtualized environments, additional load

on the system might always and unpredictably oc-

cur (noisy neighbours). This can lead to disturbed

measurement results, because performance isolation

across virtual machines might not be the case (Gupta

et al., 2006). To derive performance-relevant prop-

erties correctly, the measurements have to designed

robust w. r. t. measurement noise.

3 APPROACH

In the following, we present our approach to de-

rive performance-relevant properties automatically in

more detail.

An overview on the approach is given in Figure 1

and consists of the following steps:

1. Deploy Drivers. The approach is based on goal-

oriented measurements, which requires the target

environment to be available. On this environment,

i. e. on the virtual machines in which the target ap-

plication is to be run on, a load driver has to be

installed. For the load driver, we plan to use re-

source demands libraries which allow to generate

different patterns of CPU load and I/O load.

2. Perform Measurements. Once the driver is de-

ployed and calibrated, experiments are to be per-

formed, which include issuing different patterns

of CPU and I/O load and measuring certain parts

of the issued load. To derive performance-relevant

properties, the load patterns have to be designed

in a way that the measured results allow to infer

implementation properties of the platform through

statistical analyses. The measurements should be

robust w. r. t. measurement noise, and should be

generic, i. e. they should be applicable to miscella-

neous virtualization platforms, and not be targeted

towards a speciﬁc platform. Possible generic mea-

surements include virtualization overhead for cer-

tain resource demands, network I/O throughput

and latency rates, as well as storage access prop-

erties (for different read-write patterns). The mea-

surements are initiated by an experiment con-

troller, which also collects the measurement re-

sults from the load drivers after the measurements

have been completed.

3. Derive Performance-relevant Properties. The

measurement results serve as input for the anal-

ysis. The analyzer derives properties by using sta-

tistical methods, and triggers further experiments

if necessary.

4. Integrate Derived Properties into Analysis Tool.

Finally, the detected properties have to be inte-

grated into the analysis model. Conﬁgurable per-

formance properties can be implemented in a per-

formance analysis tool for example by using a

conﬁguration model. In this case, the detected

properties would be passed to the analysis tool as

a conﬁguration instance. Depending on the prop-

erty conﬁguration, the analysis tool has to provide

different analyzer logic in order to reﬂect the per-

formance properties.

Once the performance analysis tool is conﬁgured

based on the detected performance properties, the

software architect can conduct a performance anal-

ysis by creating a model of the application software,

which is then transformed into an analysis model. The

analysis tool uses this model to perform the analysis,

taking into account the experimentally derived perfor-

mance properties.

Our approach aims at deriving performance-

TOWARDS PERFORMANCE PREDICTION FOR CLOUD COMPUTING ENVIRONMENTS BASED ON

GOAL-ORIENTED MEASUREMENTS

619

relevant properties for a wide range of virtualized

environments. Besides virtualization environments

consisting of a single hypervisor machine running a

couple of virtual machines, the approach is also in-

tended to be applicable on larger environments, for

example cloud computing environments like Ama-

zon EC2 (Amazon.com, 2006). Note that the generic

approach could also be adapted for different prop-

erties of non-virtualized software systems, such as

performance-relevant properties of messaging sys-

tems.

To relieve the software architect of the burden of

deriving performance-relevant properties manually,

we plan to automate large parts of the approach. Both

taking the measurements and evaluating experiment

results can be done in an automated way if measure-

ments can be designed in a generic way in order to

yield robust results. If certain properties cannot be

unambiguously identiﬁed through measurements, ad-

ditional reasoning has to be done manually in order to

ﬁne-tune the performance model. To support automa-

tion, we are currently implementing a measurements

framework that is able to conduct and control the mea-

surement experiments, as well as analysing the exper-

iment results and deriving the performance-relevant

properties.

4 RELATED WORK

Design-time performance prediction has been in the

focus of research for the last years (surveyed in (Bal-

samo et al., 2004) and (Koziolek, 2010)), but virtu-

alization and cloud computing as a novel trends pro-

pose new challenges for software performance predic-

tion. An approach to model the performance impact

of server consolidation with virtual machines is pre-

sented in (Menasce and Bennani, 2006), (Menasce,

2005). Analytical queueing models are used to

provide performance estimates (i. e. response time,

throughput, utilization) of virtual machines running

on a single hypervisor. However, the system is mod-

elled at a very abstract level and only rough perfor-

mance estimates (average times only) are possible.

Other approaches provide virtualization models, but

focus only on the deployment of virtual machines or

virtual clusters (Sotomayor et al., 2006), (Yamasaki

et al., 2007). These approaches do not consider the

performance of a virtualized application.

Various approaches make use of measurements to

detect performance properties. Cherkasova and Gard-

ner (Cherkasova and Gardner, 2005) propose mea-

surements to detect a certain kind of CPU overhead

for I/O that occurs for a speciﬁc virtualization solu-

tion (Xen). They use a framework similar to the one

that is needed for our approach w. r. t. measuring and

monitoring, however the approach is only applicable

to the Xen hypervisor. Besides, the measurements

only focus on a certain scheduling property of the hy-

pervisor. Microbenchmarks are used in (Wood et al.,

2008) to estimate performance overheads that occur

when deploying an application into a virtualized en-

vironment. Different kinds of virtualization overhead

are regarded to detect resource requirements, but the

approach cannot be applied to analyze response times

of applications running in virtualized environments.

In (Woodside et al., 2001), resource functions are de-

ﬁned to derive resource demands of a software sys-

tem by using regression splines. Zheng et. al. (Zheng

et al., 2008) use Kalman ﬁlters to estimate param-

eters of the performance model. Both approaches

aim at detecting input parameters for the software

model, whereas our approach focusses on deriving

performance-relevant properties of the virtualization

environment. (Iosup et al., ress) provide initial mea-

surements of performance properties of cloud com-

puting platforms. The authors consider cloud com-

puting services for scientiﬁc computing, an integra-

tion of measurement results into analysis tools, such

as performance prediction tools, is however not in the

focus.

Challenges of measurements of MPI performance

characteristics are presented in (Gropp and Lusk,

1999). This work covers interesting issues that might

lead to incorrect measurements, but does not present

solutions for all issues. Besides, this work does not

deal with measurement challenges speciﬁc to virtu-

alization, such as inaccurate timers. Frameworks for

automated benchmarks in distributed environments or

grid environments have been proposed in (Kalibera

et al., 2006), (Tsouloupas and Dikaiakos, 2006), but

these frameworks do not cover performance proper-

ties speciﬁc to virtualization environments or cloud

computing environments. Also, an automated evalua-

tion of the benchmark results is not supported.

5 CONCLUSIONS

In this position paper, we presented a novel approach

to enable performance prediction of applications run-

ning in cloud computing and virtualized environ-

ments. The approach applies goal-oriented measure-

ments to automatically derive performance-relevant

properties of the environment. A “white-box” per-

spective on the virtualized system is not necessary,

i. e. detailed knowledge about the behaviour of vir-

tualization and cloud technologies is not required.

CLOSER 2011 - International Conference on Cloud Computing and Services Science

620

Instead, such properties are to be detected by goal-

oriented measurements on the target platform.

We are currently implementing a framework for

automated measurements in order to infer predic-

tion models of virtualised environments by systematic

measurements.

For this purpose, it has to be validated if the ap-

proach can to derive performance properties with rea-

sonable accuracy. Based on our previous experience

(Hauck et al., 2010), the approach is a promising start-

ing point for an inclusion of performance properties

of cloud computing environments or more detailed

virtualization effects into design-time software per-

formance prediction.

ACKNOWLEDGEMENTS

The work described in this paper was partially sup-

ported by the German Federal Ministry of Education

and Research (BMBF) under grant 01IC10S01A.

REFERENCES

Adams, K. and Agesen, O. (2006). A Comparison of Soft-

ware and Hardware Techniques for x86 Virtualiza-

tion. In ASPLOS-XII: Proceedings of the 12th Inter-

national Conference on Architectural Support for Pro-

gramming Languages and Operating Systems. ACM.

Amazon.com, I. (2006). Amazon EC2. Amazon Elastic

Compute Cloud. http://aws.amazon.com/ec2.

Balsamo, S., Di Marco, A., Inverardi, P., and Simeoni, M.

(2004). Model-based Performance Prediction in Soft-

ware Development: A Survey. IEEE Transactions on

Software Engineering, 30(5):295–310.

Becker, S., Koziolek, H., and Reussner, R. (2009). The

Palladio Component Model for Model-driven Perfor-

mance Prediction. Journal of Systems and Software,

82:3–22.

Cherkasova, L. and Gardner, R. (2005). Measuring CPU

Overhead for I/O Processing in the Xen Virtual Ma-

chine Monitor. In USENIX 2005: Proceedings of the

USENIX Annual Technical Conference.

Cherkasova, L., Gupta, D., and Vahdat, A. (2007). Compar-

ison of the Three CPU Schedulers in Xen. SIGMET-

RICS Performance Evaluation Review, 35(2):42–51.

Gropp, W. and Lusk, E. (1999). Reproducible Mea-

surements of MPI Performance Characteristics. In

PVM/MPI 1999: Proceedings of the 6th European

PVM/MPI Users’ Group Meeting. Springer-Verlag.

Gupta, D., Cherkasova, L., Gardner, R., and Vahdat, A.

(2006). Enforcing Performance Isolation Across Vir-

tual Machines in Xen. In Middleware 2006: Proceed-

ings of the ACM/IFIP/USENIX 2006 International

Conference on Middleware, New York, NY, USA.

Springer-Verlag.

Happe, J., Becker, S., Rathfelder, C., Friedrich, H., and

Reussner, R. H. (2010). Parametric Performance

Completions for Model-driven Performance Predic-

tion. Performance Evaluation, 67(8):694–716.

Hauck, M., Happe, J., and Reussner, R. H. (2010). Au-

tomatic Derivation of Performance Prediction Models

for Load-balancing Properties Based on Goal-oriented

Measurements. In MASCOTS 2010: Proceedings of

the 18th IEEE International Symposium on Modeling,

Analysis and Simulation of Computer and Telecom-

munication Systems. IEEE Computer Society.

Iosup, A., Ostermann, S., Yigitbasi, N., Prodan, R.,

Fahringer, T., and Epema, D. (in press). Performance

analysis of cloud computing services for many-tasks

scientiﬁc computing. IEEE Transactions on Parallel

and Distributed Systems.

Kalibera, T., Lehotsky, J., Majda, D., Repcek, B., Tom-

canyi, M., Tomecek, A., Tuma, P., and Urban, J.

(2006). Automated Benchmarking and Analysis Tool.

In VALUETOOLS 2006: Proceedings of the 1st In-

ternational Conference on Performance Evaluation

Methodolgies and Tools. ACM.

Koziolek, H. (2010). Performance Evaluation of

Component-based Software Systems: A Survey. Per-

formance Evaluation, 67(8):634–658.

Matthews, J. N., Hu, W., Hapuarachchi, M., Deshane, T.,

Dimatos, D., Hamilton, G., McCabe, M., and Owens,

J. (2007). Quantifying the Performance Isolation

Properties of Virtualization Systems. In ExpCS 2007:

Proceedings of the 2007 Workshop on Experimental

Computer Science. ACM.

Menasce, D. (2005). Virtualization: Concepts, Applica-

tions, and Performance Modeling. In CMG 2005:

Proceedings of the International CMG Conference.

Menasce, D. and Bennani, M. (2006). Autonomic Virtu-

alized Environments. In ICAS 2006: Proceedings of

the 2nd International Conference on Autonomic and

Autonomous Systems.

Schroeder, B., Wierman, A., and Harchol-Balter, M. (2006).

Open Versus Closed: A Cautionary Tale. In NSDI

2006: Proceedings of the 3rd Conference on Net-

worked Systems Design & Implementation. USENIX

Association.

Sotomayor, B., Keahey, K., and Foster, I. (2006). Overhead

Matters: A Model for Virtual Resource Management.

In VTDC 2006: Proceedings of the 2nd International

Workshop on Virtualization Technology in Distributed

Computing.

Tsouloupas, G. and Dikaiakos, M. D. (2006). Characteri-

zation of Computational Grid Resources Using Low-

Level Benchmarks. In E-SCIENCE 2006: Proceed-

ings of the 2nd IEEE International Conference on e-

Science and Grid Computing. IEEE Computer Soci-

ety.

Wood, T., Cherkasova, L., Ozonat, K., and Shenoy, P.

(2008). Proﬁling and Modeling Resource Usage of

Virtualized Applications. In Middleware 2008: Pro-

ceedings of the 9th ACM/IFIP/USENIX International

Conference on Middleware. Springer-Verlag.

TOWARDS PERFORMANCE PREDICTION FOR CLOUD COMPUTING ENVIRONMENTS BASED ON

GOAL-ORIENTED MEASUREMENTS

621

Woodside, C. M., Vetland, V., Courtois, M., and Bayarov, S.

(2001). Resource Function Capture for Performance

Aspects of Software Components and Sub-Systems.

In Performance Engineering, State of the Art and Cur-

rent Trends, pages 239–256. Springer-Verlag.

Yamasaki, S., Maruyama, N., and Matsuoka, S. (2007).

Model-based Resource Selection for Efﬁcient Virtual

Cluster Deployment. In VTDC 2007: Proceedings

of the 3rd International Workshop on Virtualization

Technology in Distributed Computing.

Zheng, T., Woodside, C. M., and Litoiu, M. (2008). Perfor-

mance Model Estimation and Tracking Using Optimal

Filters. IEEE Transactions of Software Engineering,

34(3):391–406.

CLOSER 2011 - International Conference on Cloud Computing and Services Science

622