PERFORMANCE OVERHEAD OF ERP SYSTEMS IN

PARAVIRTUALIZED ENVIRONMENTS

André Bögelsack

Department of Informatics, Technische Universitaet Muenchen

Boltzmannstraße 2, 85748 Garching, Germany

Keywords: Performance, ERP system, Virtualization, Zachmann test.

Abstract: Virtualization is a big trend in current IT world and is used intensively in today’s computing centres. But

little is known about what happens to the performance of computer systems when running in virtual

environments. This work focuses on the performance aspect especially in the field of Enterprise Resource

Planning systems (ERP). Therefore, this work utilizes a quantitative approach by using laboratory

experiments to measure the performance differences between a virtualized and non-virtualized ERP system.

First, on basis of a literature review a performance measurement framework will be developed to provide a

comprehensive guideline how to measure the performance of an ERP system in a virtualized environment.

Second, the performance measurement focuses on the overhead in CPU, memory and I/O intensive

situations. Third, the focus lays on a root cause analysis. Gained results will be analyzed and interpreted to

give recommendations for further development of both ERP system and virtualization solution. The

outcome may be useful for further computing centre design when introducing new ERP systems and service

delivery concepts like Software as a Service (SaaS) in a virtualized or non-virtualized environment.

1 INTRODUCTION

Virtualization’s history starts in 1964 with the

research project called CP-40 from IBM (Donovan

and Madnick 1975). Goal of this project was the

development of the so called virtual machine/virtual

memory time-sharing operating system. Research

output from CP-40 was utilized in the IBM

System/360-67 (Donovan and Madnick 1975). This

mainframe was equipped with an operating system

which was able to share the resource for several

users and run several instances of an operating

system in so called virtual machines. Later in 1974

Popek and Goldberg (Popek and Goldberg, 1974)

defined the formal requirements for virtualization,

which are valid until today and form the main

guidelines for all developers of virtualization

solutions. During the 1970’s till the 1990’s

virtualization was well established in the world of

mainframes, but not in the world of conventional

x86 personal computers. This is because of the x86

architecture was not built to support virtualization

and does not meet the requirements for virtualization

(Robin and Irvine, 2000). A first virtualization

solution for x86 world was launched by VMware in

1999 (Devine et al., 1998). This solution overcomes

the gap between the x86 architecture and the

virtualization idea by bringing in a technique called

trapping. Trapping is very time and CPU consuming

because every executed instruction in the virtual

machines has to be scanned during runtime

(LeVasseur et al., 2006). Utilizing trapping in a

virtualization solution assigns a virtualization

solution to the type of “full-virtualization”.

In 2001 a project called Denali was launched,

which followed the idea of lightweight virtual

machines (Whitaker et al., 2002). They developed a

kind of virtualization solution for running specially

adopted, slim operating systems in their virtual

machines. Their main goal was the adoption of an

isolation kernel for performance, scale and

simplicity. This is often cited as the beginning of

“paravirtualization” which may be described as

second type of virtualization (Nakajima and Mallick,

2007). In 2003 a development project called Xen

from the university of Cambridge used another

method to run virtual machine on x86 hardware

(Barham et al., 2003). They moved from the

trapping idea to the idea of modifying guest

operating systems to be aware of an underlying

virtual machine monitor. This leads to a

performance improvement compared to full-

200

Bögelsack A. (2009).

PERFORMANCE OVERHEAD OF ERP SYSTEMS IN PARAVIRTUALIZED ENVIRONMENTS.

In Proceedings of the 11th International Conference on Enterprise Information Systems, pages 201-207

DOI: 10.5220/0002194102010207

 SciTePress

virtualization but has the disadvantage that every

guest OS has to be modified before running on top

of Xen. The main motivation for Xen was to

improve performance of virtualization.

When dealing with the performance of

virtualization solutions there are three main aspects

to pay attention to: CPU, I/O and memory (Huang et

al., 2006). Every guest OS operation accessing CPU,

I/O or memory produces an overhead because of the

work of the hypervisor. The hypervisor’s task is to

intercept these accesses. Depending on the type of

virtualization this may produce more (full-

virtualization) or less overhead (paravirtualization).

A lot of research work has been done for the Xen

hypervisor dealing with overhead measurements for

file accesses, network accesses or CPU accesses (see

Related Work). But to our knowledge there are no

performance overhead measurements dealing with

the performance of ERP systems when running in

virtual machines. In (Casazza et al., 2006) the need

for a new kind of virtualization benchmark is

discussed because existing benchmarks may not be

adapted to virtual environments. The main argument

is that isolated benchmarks may not be suitable to

cover all performance topics in virtual environments.

This work follows the idea and argues that ERP

systems show different performance behaviour than

isolated benchmarks show. As ERP systems build

the backbone of today’s most important core

business processes, nobody knows what happens

when running an ERP system in a virtualized

environment. There are no scientific publications

available. So, this research work focuses on the

performance overhead for ERP systems running in

paravirtualized environments.

2 RESEARCH GOALS

To evaluate the impact of paravirtualization on the

performance of ERP system, this work follows the

central hypothesis: paravirtualization influences

ERP system’s performance negatively. Such

hypothesis may be derived from former work in the

area of performance overhead research which

attested e.g. a performance loss up to 50% for I/O

processing in a virtualized environment (Cherkasova

and Gardner, 2005). Adopting these results to the

area of ERP systems gives a starting point for this

work as ERP systems are I/O intensive and may

experience performance degradation too.

This very general hypothesis is elaborated into

three main research questions, which are oriented

after Barham (Barham et al., 2003), Huang (Huang

et al., 2006), Svobodova (Svobodova, 1976) and Jain

(Jain, 1991).

1. How can the ERP system’s performance be

measured in a paravirtualized environment and

which performance measures are important?

The purpose of dealing with the question of how

to measure performance in this setting is to find out

how the ERP system can be stressed so that its

operations produce a massive workload on the

hypervisor and how the operations of the ERP

system as well as the hypervisor can be monitored.

The work focuses on paravirtualized environments

as such environments are known for a better

performance than full-virtualized environments

(Nakajima and Mallick, 2007). By analyzing

available literature, appropriate benchmark tools will

be discovered which are suitable for the ERP system

and are able to produce massive workload on the

ERP system. The range of benchmark tools will be

limited by the type of workload the tools provide

and to which ERP system the benchmark tools can

be applied to. This research question does also cover

the discovery of appropriate performance measures

whereas performance measures can be understood as

“a proper set of parameters upon which the

evaluation will be based” (Svobodova, 1976). The

set of performance measures is limited by the

benchmark tool as well as the ERP system itself.

The choice which performance measures are

necessary or not is oriented after the three overhead

aspects CPU, memory and I/O (Huang et al., 2006).

After determining the needed performance measures

the work focuses on the available performance

monitors. Regardless if these monitors are hardware

or software monitors they must be able to return the

necessary numerical performance data. The number

of performance monitors will be limited by the used

hypervisor, ERP system, benchmark tool,

performance measures and operating system.

Finally, within the scope of this research question

we research possible testing environments

(breadboard construction) which include test

hardware, test software, possible workloads and

performance measures.

By utilizing a literature review for answering the

first research question the outcome will include

possible benchmarks for the ERP system,

performance measures and an appropriate testing

environment.

2. Does a hypervisor significantly influence

CPU, memory and I/O operations of an ERP system

working in a paravirtualized environment in high

workload situations?

Stressing the ERP system with distinct CPU,

PERFORMANCE OVERHEAD OF ERP SYSTEMS IN PARAVIRTUALIZED ENVIRONMENTS

201

memory and I/O intensive workloads we assume that

the hypervisor will be stressed too because of the

previously mentioned interceptions. Although there

are some publications available dealing with CPU

and memory overhead we believe due to the

architecture of the ERP system (with a rather large

number of concurrently active processes) we will get

different results. The work focuses on high workload

situations as these situations influence the subjective

user experience when working with an ERP system

e.g. due higher response times. To quantify the

assumed performance degradation we will run

several test runs inside the paravirtualized

environment and compare them to results gathered

in a native environment. Modifying the

configuration of the hypervisor (e.g. scheduler) and

ERP system allows us to determine which

configuration suits best the interaction of the ERP

system and the hypervisor. The hypervisor’s impact

on performance can be derived already from the

architecture and previous work (Barham et al., 2003)

but the answer should be deduced in an empirical

way to be able to quantify the impact.

Answering the second research question enables

us to highlight the performance impact of

paravirtualization on massive CPU, memory and I/O

loads when running an ERP system.

3. What causes the performance overhead of the

ERP system in a paravirtualized environment?

The last research question focuses on the root

cause analysis of performance degradation. On basis

of extensive monitoring inside the hypervisor, the

ERP system and the operating system during the test

runs we aim on statements what causes the overhead

(e.g. swapping problems, scheduler problems, lock

problem inside ERP system etc). Analysis’s findings

may also be of interest for other hypervisors.

Cherkasova (Cherkasova et al., 2007) tested several

configurations for the internal scheduler in Xen and

pointed out performance differences due to different

scheduler configurations. By utilizing monitors to

gather data and by changing the configuration during

the test runs we can analyze system behavior in

detail.

Researching the aforementioned topics the work

as a whole highlights the performance impact of

paravirtualization on an ERP system when running

massive CPU, memory and I/O operations. In

addition it allows for recommendation for choosing

the best ERP system configuration and gives insights

into the reason for performance degradation.

3 LIMITATIONS

This work is limited in its scope by two factors.

First, the work is limited to a specific virtualization

solution. Second, this work chooses a specific ERP

system.

The current market for virtualization is very

large: several different virtualization solutions form

a heterogeneous market (Jehle et al., 2008). On top

of the heterogeneity there are several virtualization

techniques and kinds of virtualization available

(Nakajima and Mallick, 2007). Because of the better

performance of paravirtualization this work chooses

paravirtualization as a field of focus for the

performance tests. Currently the main research focus

lays on Xen, which is a typical paravirtualization

solution. As Xen is open-source, several vendors

began to develop their own implementation of this

hypervisor. In this work one of these

implementations from Sun Microsystems™ is used.

Another limitation is the ERP system. The

current market in the field of ERP system is

dominated by SAP. Most of the Fortune 500

companies use SAP software to support their core

business processes. As SAP is the market leader and

this work is positioned in the SAP University

Competence Centre, the used ERP system for

evaluation will be SAP ERP 2005. A slightly

modified edition of this ERP system is available as

the so called SAP Linux Certification Suite (SLCS)

(Kühnemund, 2007). It is used internally by SAP to

test and benchmark new hardware platforms for the

first time. As the server hardware, which will be

used for the performance evaluation in this work, is

provided by Sun Microsystem™, a specially ported

version for this hardware platform will be used.

4 RELATED WORK

The current research focus lies on paravirtualization.

There are several works available dealing with the

performance impact of paravirtualization in different

ways and in different fields of applications ((Huang

et al., 2006), (Cherkasova and Gardner, 2005) ,

(Mennon et al., 2005), (Zhang and Dong, 2008),

(Youseff et al., 2006)). For the performance of ERP

system there are many works available dealing with

the user’s performance in ERP system but only a

few works focus on the performance of the ERP

system itself ((Kemper et al., 1999), (Wilhelm,

2003), (Zeller and Kemper, 2002)). Therefore the

current research status is divided into two subsets: a

first subset describes related work in the area of

Special Session DR 2009 - Special Session on Doctoral Research

202

performance measurement in paravirtualized

environments. A second subset explains the current

work in the area of ERP system’s performance.

4.1 Performance Measurements in

Paravirtualized Environments

The first work to be pointed out is the work of

Barham et al. (Barham et al., 2003), who developed

Xen, a hypervisor operating with a new concept for

virtualization. To testify their newly developed

hypervisor they ran several test runs dealing with

some typical fields of applications: SPEC INT2000

benchmark, building a Linux kernel, Open Source

Database Benchmark suite, file system benchmark

and a SPEC Web99 benchmark. In addition Barham

ran several smaller operating system benchmarks for

examining the performance overhead in particular

subsystems. Several other test runs focus on the

performance isolation capability of Xen. The result

of the work is a small performance loss in every test

run compared to a native environment. As

virtualization represents an additional software layer

this overhead is intuitively understandable.

In 2005 Cherkasova et al. (Cherkasova and

Gardner, 2005) measured the CPU overhead for I/O

processing in Xen. To do so they utilized a

SPECweb’96 and SPECweb’99 benchmark to stress

the CPU with massive I/O and network traffic. They

discovered that heavy I/O and network traffic can

produce a huge CPU overhead in the control domain

of Xen. In 2007 Cherkasova made suggestions how

to improve the performance by applying CPU

schedulers (Cherkasova et al., 2007).

Huang performed a NAS parallel benchmark to

get a first impression of the overhead which is

produced by using virtualization (Huang et al.,

2006). The results showed performance degradation

of up to 12% - 17%. Actually they looked for a new

bypass I/O which allows virtual machines to access

I/O devices directly.

Mennon focused on the performance overhead

for network traffic in 2005 (Mennon et al., 2005).

Therefore, a tool named Httperf was used and they

point out a performance overhead up to 25%

depending on the network traffic actions.

Zhang and Dong investigated the performance

overhead when using the new hardware-supported

virtualization technique from Intel, called Intel VT

(Zhang and Dong, 2008). The paper states that a

performance overhead between 5% - 20% can be

experienced. To gain these results they utilized

several benchmarks like Kernel build, SPEC Int or

SPEC JBB.

Youseff et al. compared several HPC (high

performance computing) kernels to the Xen kernel

(Youseff et al., 2006). The authors have done a

performance evaluation by using several different

benchmarks for e.g. disk performance,

communication or memory performance. Their

publication shows a performance loss up to 30%

(high workload) depending on the field of

application.

In the area of HPC a lot of work was done

regarding HPC application performance in

virtualized environments ((Ranadive et al., 2008),

(Farber, 2006), (Liu et al., 2006)). They do not deal

with the overhead issue but the overall performance.

As an essence one can derive that there is a lot of

work available regarding the measurement of

performance and performance overhead when using

Xen. Each work uses well established benchmarks.

But there is no work available which deals with the

performance overhead when running an ERP system

in a virtualized environment.

4.2 Performance Measurements of

ERP systems

Compared to the available publications for the

performance of virtualized environment, current

work for the measurement of an ERP system’s

performance are rare. This work chooses SAP ERP

2005 as the market leader to be an instance of such

ERP software systems. For SAP ERP 2005 there are

three significant works.

Kemper discussed the possibility of tuning the

performance for SAP R/3 systems (Kemper et al.,

1999). They focused on the main memory

management and the database. For the database they

mentioned several available benchmarks, like TPC-

D. In another work Zeller and Kemper (Zeller and

Kemper, 2002) showed how to increase the

performance for the database in case of special

database operations. They especially paid attention

to the main architecture of SAP R/3 by using a

special developed testing tool called SSQJ but

mainly focus on the database and not on the entire

ERP system.

In his PhD dissertation Kai Wilhelm discusses

storage systems serving a SAP R/3 system

(Wilhelm, 2003). The focus lies on characterization

of the storage load during benchmarks, the

performance measurement as well as the

recommendation for the ‘right’ storage system in a

future SAP system landscape.

In summary, former research work focussed on

the database and the underlying storage system

whereas an improvement of the overall performance

PERFORMANCE OVERHEAD OF ERP SYSTEMS IN PARAVIRTUALIZED ENVIRONMENTS

203

of the SAP R/3 system was achieved by very special

improvements in one of the components of the SAP

R/3 system. Keeping this in mind and connecting

this with the research which was done in the field of

virtualization there is a need for an evaluation of the

performance of SAP R/3 systems in virtualized

environments.

5 RESEARCH STRUCTURE

The research structure in this work is influenced by

quantitative research design and induction (Bortz

and Döring, 2005 ). The research strategy in this

work will be in accordance to the work of

(Svobodova, 1976) and (Jain, 1991). In Svobodova’s

book about computer performance measurement she

introduced three main steps for performance

measurement:

1. Define performance measures.

2. Determine the quantitative value of

performance measures and analyze system

performance with respect to system

structure and system workload.

3. Assign qualitative values to different levels

of performance measures and assess

systems performance.

Jain independently developed a more detailed

view on the performance measurement process and

proposed ten steps. Jain’s ten steps together with

Svobodova’s three steps form a comprehensive

performance measurement process. Mapping these

steps onto the described research questions draws

the research process for this work which is oriented

after Jain’s ten steps and goes into more detail with

Svobodova’s steps where it is necessary and

supportive.

1. The goal of this performance measurement

work is to find out if ERP system’s

performance differs in virtualized

environments compared to non-virtualized

environments. This work focus on ERP

system’s performance as little is known in this

field of application. Paravirtualization is

chosen from the types of virtualization as the

most research work is done in the filed of

paravirtualization. The system will consist of

an exemplary ERP system from SAP, which is

running on hardware from Sun Microsystem.

2. System services in an ERP system are

manifold. This work focus on the ability of the

ERP system to allocate memory and work in

the main memory of the underlying

hardware/operating system. This ability is

needed by the system to process user requests.

The faster system works in the memory the

faster it can respond to user requests. Another

system service is the ability to store data

persistently in a database. Therefore the system

utilizes the underlying hardware/operating

system, too. Paravirtualization may influence

the ability and the performance of the ERP

system. Hence, we need an appropriate

workload which is able to stress the ERP

system with such memory and I/O operations.

Stressing the ERP system with high memory

load will automatically stress the CPU, too. A

possible outcome may be a number of the

highest value for memory and I/O operations of

the ERP system.

3. The performance metrics are the criteria used

to compare the performance and they are

derived from the system, the workload and the

architecture which was chosen to be measured.

In this research work one can think of simple

metric for the memory operations as well as a

simple metric for the I/O operations, e.g.

throughput. At this stage some thoughts of

Svobodova can be added as she mentioned the

so called performance measures. The term

“performance measures” can be understood as

a set of parameters upon which the evaluation

will be based. Keeping this in mind and

remembering the first research question in this

work, the first step in the research process is to

determine the appropriate performance

parameters. This includes a literature review to

find out which performance parameters are

suitable and how they can be measured. Of

course performance parameters depend on the

system architecture. The literature review is

oriented after the system architecture and

covers the main components as well as their

performance parameters.

4. Fourth step is about listing all system and

workload parameters, which affect the

performance. System parameters do not change

during the measurements and workload

parameters change during the measurement.

This means parameters like the basic

information about hardware, configuration of

operating system and configuration of the ERP

system itself will be kept constant. A workload

Special Session DR 2009 - Special Session on Doctoral Research

204

parameter is the usage of a virtualization

solution. After a first test run the list of

parameters may be extended.

5. During the test runs a lot of factors are to be

studied. Factors are parameters that change

during a test run. Such factors may be slight

changes to the system’s settings but also bigger

changes like the usage of a virtualization

solution (or not). This work will start with a

short list of parameters, which then will be

extended if necessary. Parameter’s list depend

on the determined performance metrics (see

step 3) as well as system outcomes (see step 2).

6. Selecting the evaluation method involves the

measurement of the system. Besides simulation

and analytical modelling, measuring is one of

the possible evaluation methods. This work

uses measurement because the measurement

subject is a system which already exists.

7. Determination of the correct system workload

is an outcome of research question 1. The most

important step is to “characterize the workload

by distributions of demands made on individual

system resources” as well as to “define a unit

of work and express the workload as a number

of such units” (Svobodova, 1976) For the

measurement this work utilizes as synthetic

workload, which can be described as an

“artificial reproducible workload”. Task of the

workload is to produce heavy memory and

CPU load as well as heavy I/O load. Several

workloads are available for SAP systems (SAP

Benchmarks.,01/20/2009). Selection of a

suitable workload is made by answering

research question 1.

8. Designing the experiment means to establish

the laboratory environment as well as the

breadboard construction. After establishing the

environment several test runs will be run. That

will lead to the determination of the

quantitative value and allow for analyzing the

systems’ performance. This step is the core one

in the work as it involves several measurements

of the ERP system in non-virtualized

environment and paravirtualized environment.

For estimating the performance this work uses

a combination of measurements and analysis.

For the analysis of the system’s performance

several test runs must be run to gain the

necessary data. Barham (Barham et al., 2003)

used seven test runs for each test. This work

will follow this idea but will add two more test

runs as a warm-up phase for the ERP system.

This is important because the ERP system has

to fill its internal buffers to be fully operational.

Filling the buffers in a first test run may falsify

the test results. For recording the performance

parameters it is necessary to utilize

performance monitors. Performance monitor

are “tools that facilitate analytic

measurements” (Svobodova, 1976) and

“observe the activities on the system” (Jain,

1991). Generally there are two types of a

monitor: hardware and software monitors

(Leung,1988).

9. Analysing and interpreting the data is

described as translating the quantitative values

into qualitative values. This means to find

some scale of ‘goodness’ for the relative

measure of non-virtualized ERP system and

paravirtualized ERP system. The results of this

step are qualitative statements about the

performance difference of ERP systems

running in a –non-virtualized environment

compared to paravirtualized environment.

10. The last step is about presenting the results

which means to prove or reject the hypothesis.

The following table gives an overview about the

mapping of the research question to this iterative

process:

Table 1: Mapping Jain's steps onto research questions.

Research

Question Step (after (Jain, 1991))

Motivation

Goals and system boundaries

System services and outcomes

Performance metrics

System and workload parameters

Performance factors

Evaluation techniques

Workload selection

Experiments design

2 & 3

Analyze and interpret data

Present results

To outline the research strategy in few words: this work is

hypothesis-driven (as the work assumes a performance

overhead caused by paravirtualization), it uses laboratory

experiments for performance measurement to determine

quantitative values. The work utilizes a synthetic workload

and uses quantitative analysis for interpreting the gained

results.

PERFORMANCE OVERHEAD OF ERP SYSTEMS IN PARAVIRTUALIZED ENVIRONMENTS

205

6 PRACTICAL IMPACT

The most significant outcome will be the

performance results. As there are no scientific

publications about the measurement of

paravirtualized ERP systems, it is an addition to

research in this field to have an evaluation of how

these complex software systems behave in

virtualized environments. This may have an impact

on the future computer centre design when new ERP

systems are introduced to the computing centre (with

or without virtualization).

Another outcome of this work will include

recommendations for the further development of

hypervisors and ERP systems. As the work will

analyse the virtualized ERP system’s performance it

can state out possible bottlenecks and may suggest

changes to the design of hypervisor/ERP system as

well as the interaction between both components.

7 TIMELINE

Table 2: Timeline.

REFERENCES

Donovan, J. J., Madnick, S. E., Hierarchical approach to

computer system integrity. In: IBM Systems Journal,

Volume 14, No. 2, Pages 188-202 (1975)

Popek, G., J., Goldberg, R., P., Formal Requirements for

Virtualizable third Generation Architectures. In:

Communications of the ACM, Volume 17, Number 7,

Pages 412-421 (1974)

Robin, J., S., Irvine, C., E., Analysis of the Intel Pentium’s

Ability to Support a Secure Virtual Machine Monitor.

In: SSYM'00: Proceedings of the 9th conference on

USENIX Security Symposium (2000)

VMWARE Devine, Scott W., Bugnion, Edouard,

Rosenblum, Mendel, Virtualization system including a

virtual machine monitor for a computer with a

segmented architecture (1998)

LeVasseur, J., Uhlig, V., Chapman, M., Chubb, P., Leslie,

B., Heiser, G., Pre-Virtualization: soft layering for

virtual machines. Technical Report 2006-15, TH

Karlsruhe (2006)

Whitaker, A., Shaw, M., Gribble, S. D., Scale and

performance in the Denali isolation kernel. In:

Proceedings of the 5

symposium on Operating system

design and implementation, Pages 195-209 (2002)

Nakajima, J., Mallick, A. K., Hybrid-Virtualization –

Enhanced Virtualization for Linux. In: Linux

Symposium, Volume 2, Pages 87-96 (2007)

Barham, P., Dragovic, B., Fraser, K., Hand, S., Harris, T.,

Ho, A., Neugebauer, R., Pratt, I., Warfield, A., Xen

and the Art of Virtualization. In: SOSP ’03 -

Proceeding of the nineteenth ACM symposium on

Operating systems principles, Pages 164-177, (2003)

Huang, W., Jiuxing, L., Abali, B., Dhabaleswar, K. P., A

Case for High Performance Computing with Virtual

Machines. In: ICS ’06: Proceedings of the 20

annual

international conference on Supercomputing, Pages

125-134 (2006)

Casazza, J. P., Greenfiled, M., Shi, K., Redefining Server

Performance Characterization for Virtualization

Benchmarking. In: Intel Technology Journal, Volume

10, Issue 3, Pages 243-251 (2006)

Cherkasova, L., Gardner, R., Measuring CPU Overhead

for I/O Processing in the Xen Virtual Machine

Monitor. In: Proceeding of the annual conference on

USENIX Annual Technical Conference, Pages: 387-

390 (2005)

Svobodova, L., Computer Performance Measurement and

Evaluation Methods: Analysis and Applications.

American Elsevier Publishing Company, Inc., New

York (1976)

Jain, R., The art of computer systems performance

analysis: techniques for experimental design,

measurement, simulation, and modelling. John Wiley

& sons, Inc., Littleton, Massachusetts (1991)

Cherkasova, L. Gupta, D., Vahdat, A., Comparison of the

Three CPU Schedulers in Xen. In: ACM SIGMETRICS

Performance Evaluation Review, Volume 35, Issue 2,

Pages 42-51 (2007)

Jehle, H., Wittges, H., Bögelsack, A., Krcmar, H.:

Virtualisierungsarchitekturen für den Betrieb von Very

Large Business Applications. In: Proceedings of

Multikonferenz Wirtschaftsinformatik, Pages 1901-

1912 (2008)

Kühnemund, H., Documentation for SLCS v2.3. SAP AG,

Walldorf (2007)

Mennon, A., Santos, J. R., turner, Y., Janakiraman, G. J.,

Zwaenepoel, W, Diagnosing performance overheads

in the Xen virtual machine environment. In:

Proceedings of the 1

ACM/USENIX international

conference on Virtual execution environments, Pages

13-23 (2005)

Zhang, X., Dong, Y., Optimizing Xen VMM Based on

Intel® Virtualization Technology. In: Proceedings of

the 2008 International Conference on Internet

Computing in Science and Engineering, Pages 267-

274 (2008)

Youseff L., Wolski R., Gorda B., Krintz C., Evaluating the

Performance Impact of Xen on MPI and Process

Execution For HPC Systems. In: Proceedings of

Duration /

Milestone

Activity

01/2009 –03/2009 Literature review for RQ 1

03/2009 –12/2009 Performance measurements

and test runs for RQ 2

12/2009 –06/2010 Root cause analysis for RQ 3

07/31/2010 Dissertation version 1.0

Special Session DR 2009 - Special Session on Doctoral Research

206

Second International Workshop on Virtualization

Technology in Distributed Computing (2006)

Kemper, A., Kossmann, D., Zeller, B., Performance

Tuning for SAP R/3. In: Data Engineering, Volume

22, Number 2, Pages 32-39 (1999)

Wilhelm, K., Messung und Modellierung von SAP R/3-

und Storage-Systemen für die Kapazitätsplanung

(2003)

Zeller, B., Kemper, A., Exploiting Advanced Database

Optimization Features for Large-Scale SAP R/3

Installations. In: Proceedings of the 28

VLDB

Conference (2002)

Ranadive, A., Kesavan, M., Gavrilovska, A., Schwan, K.,

Performance Implications of Virtualizing Multicore

Cluster Machines. In: Proceedings of the 2

Workshop on System-level Virtualization for High

Performance Computing (2008)

Farber, R., Keeping “Performance” in HPC: A look at the

impact of virtualization and many-core processors. In:

Scientific Computing (2006)

Liu, J., Huang, W., Abali, B., Panda, D., K., High

Performance VMM-Bypass I/O in Virtual Machines.

In: ATC (2006)

Bortz, J. and N. Döring (2005): Forschungsmethoden und

Evaluation für Human- und Sozialwissenschaftler. 3,

Springer, Heidelberg 2005.

SAP Benchmarks. http://www.sap.com/benchmarks,

accessed on 01/20/2009

Leung, C. H. C., Quantitative Analysis of Computer

Systems. John Wiley & Sons, Inc., London (1988)

PERFORMANCE OVERHEAD OF ERP SYSTEMS IN PARAVIRTUALIZED ENVIRONMENTS

207