Grid resource management with service level agreements [Elektronische Ressource] / Tianchao Li

Sujets

• Informatik

Informations

• Publié par: technische_universitat_munchen
• Publié le: 01 janvier 2008
• Nombre de lectures: 13
• Langue: English
• Poids de l'ouvrage: 3 Mo

Extrait

Lehrstuhl für Rechnertechnik und
Rechnerorganisation
der Technischen Universität München
Grid Resource Management with Service Level
Agreements
TianchaoLi
Vollständiger Abdruck der von der Fakultät für Informatik der Technischen Universität
München zur Erlangung des akademischen Grades eines
Doktors der Naturwissenschaften (Dr. rer. nat.)
genehmigten Dissertation.
Vorsitzender: Univ.-Prof. Dr. A. Bode
Prüfer der Dissertation:
1. Univ.-Prof. Dr. H. M. Gerndt
2. Univ.-Prof. Dr. M. Bichler
Die Dissertation wurde am 21.01.2008 bei der Technischen Universität München eingere-
icht und durch die Fakultät für Informatik am 02.07.2008 angenommen.Abstract
With Grid computing adopting Service Oriented Architecture (SOA), resources are repre-
sented by services with standardized interfaces and the management of converges
to the more general management of services (i.e. the Service Level Management). The key
to Service Level Management is the management of Service Level Agreement (SLA), a for-
mal contract between the provider and consumer of a service which stipulates and commits
the provided service to a required level of service.
We concentrate in our research on designing and establishing an infrastructure for SLA
management. It consists of a site-local infrastructure for managing SLAs and local re-
sources, and a global infrastructure for advertising and brokering of resources.
From the client’s point of view, a typical session of service usage includes the following
activities: the client sends job parameters and SLA parameters (such as deadline, price
etc.) to the broker, the broker contacts the service sites to get available SLA templates,
the broker selects the appropriate template and sends it to the client for verification and
an acknowledgement is represented by the return of a signed SLA offer. The SLA offer is
then forwarded by the broker to the corresponding service site to create the SLA. With the
negotiated SLA, the client invokes the service and gets an id (endpoint reference) of the
job. According to the negotiated SLA, the job is scheduled at appropriate time and as soon
as the job finishes, a notification is sent to the client so that the result can be retrieved.
On the service site, a service has an associated management infrastructure that handles
the management of SLAs and resources. The WS-Agreement services provide interfaces
for requests for SLA templates, SLA creation and termination, notification of SLA status
changes etc. The performance predictor estimates resource demands for specific values
of job parameters, and the resource manager keeps track of resource capabilities and job
schedules. Based on the information provided by both, the agreement decision maker de-
cides on the acceptance or rejection of a specific SLA offer. The resource manager (together
with the underlying scheduler) is also responsible for the scheduling of jobs with resource
reservation and service invocation at scheduled time.
Targeting to the SLA-based resource management infrastructure, we focus on the fun-
damental issues for the establishment of such an including the specification
of SLAs with WS-Agreement and the design and implementation of a generic and ex-
tensible support infrastructure for WS-Agreement specification, the evaluation of various
techniques for application performance prediction and the establishment of a generic in-
frastructure for non-intrusive monitoring and adaptive prediction of the execution time of
iservices, discussion and solution of job scheduling issues in the local resource management
infrastructure with a focus on the cases where services have exclusive access to the local
resource. In addition, we also present the design and implementation of a set of develop-
ment tools and support environments for the development, management, and accessing of
service-oriented Grid applications that are integrated with our SLA-based resource man-
agement infrastructure.
iiAcknowledgments
I would like hereby to express my gratitude to all that have made suggestions and offered
help towards the successful completion of this work, and to all that have supported me from
different aspects during the past years.
First of all I would like to thank my advisor, professor Michael Gerndt. He not only
directed me to the right way for pursuing my work but also gave me the most considerate
help throughout the whole course of this research work. I would like to express my special
gratitude to professor Arndt Bode, the leader of our research chair, for all his support and
encouragement. I am also grateful to him for his help in overcoming all the financial and
administrative issues. I would also like to thank professor Martin Bichler for taking the
time to review this thesis and giving valuable comments.
I want to express my gratitude to all my colleagues and the technical and administrative
staff at LRR-TUM. I would like to specially mention Dr. Jie Tao, Dr. Edmond Kereku, Dr.
Josef Weidendorfer, and Houssam Haitof, who worked together with me on projects, and
Dr. Daniel Stodden, who started and finished the Ph.D. study at about the same time and
discussed many issues towards successful completion of the Ph.D.
My research work documented in this thesis has been supported by IBM Center of Ad-
vanced Studies with a research grant for Autonomous Resource Management for Large
Scale Applications. Thanks to Dr. Toni Bollinger, Dr. Hans-Dieter Wehle, and Dr. Niko-
laus Breuer from IBM Development Laboratory in Boeblingen for all the valuable discus-
sions and help in both the application of research funding from IBM and the whole course
of this project.
Finally, thanks to my wife who accompanied me and helped me taking care of many
issues so that I can concentrate on the work. Thanks to my son, Tony, who has brought
happiness, sorrow, hope, and a lot more to me. You have helped me to understand the true
meaning of life.
Tianchao Li
Munich, Germany
December 2007
iiiContents
1 Introduction 1
1.1 History and Status 1
1.2 Methodology and Outline 2
2 An Infrastructure for SLA-based Resource Management 5
2.1 Introduction 5
2.2 An Infrastructure for SLA-based Resource 5
2.3 Challenges 6
2.4 Application Scenarios 8
2.5 Summary 8
3 Specification and Management of Service Level Agreement 9
3.1 Introduction 9
3.2 WS-Agreement Specification 10
3.2.1 Overview 10
3.2.2 Agreement Schema 11
3.2.3 Template Schema 12
3.2.4 PortType Definition 12
3.2.5 Involvement and Contributions 14
3.3 WS-Agreement Support Infrastructure 14
3.3.1 Generic Design 14
3.3.2 WS-Agreement Services 14
3.3.3 Agreement Management and Monitoring 15
3.3.4 Domain Specific Handling with Backend Providers 17
3.4 Implementation Issues 19
3.4.1 GT4 Platform 19
3.4.2 Interface Schema Adaptation 20
3.4.3 Domain Specific Terms 20
3.4.4 Provider Management 21
3.4.5 Agreement Monitoring 21
3.4.6 Security Issues 22
3.5 Related Work 22
i3.5.1 SLA Languages 22
3.5.2 SLA Infrastructures 22
3.5.3 WS-Agreement Implementations 23
3.6 Experimental Results 25
3.6.1 Performances 25
3.6.2 Examples 25
3.7 Conclusion 25
4 Prediction-based Application Performance Evaluation 27
4.1 Introduction 27
4.2 The Need for Performance Prediction 28
4.3 Feasible Application Performance Prediction Techniques 29
4.3.1 Analytical Model 29
4.3.2 Statistical Simulation 29
4.3.3 Historical Data Analysis 30
4.4 A Systematic Approach for Performance Prediction with Data Mining
Techniques 32
4.4.1 Overview 32
4.4.2 Performance Data Collection 33
4.4.3 Data Mining 33
4.4.4 Data and Model Management 34
4.5 A Run Time Monitoring and Prediction Framework for Grid Services 34
4.5.1 Overview 34
4.5.2 Non-intrusive Data Collection 35
4.5.3 Generic Modeling and Prediction Interface 36
4.5.4 Automatic Management of Performance Model 36
4.6 Implementation Issues 36
4.6.1 The Globus SOAP Handler 36
4.6.2 Performance Data Storage 38
4.6.3 Prediction 39
4.6.4 Configuration and Management of Recorders and Predictors 39
4.7 Existing Work on Application Performance Prediction 40
4.7.1 Analytical Model 40
4.7.2 Statistical Simulation 40
4.7.3 Historical Data Analysis 42
4.8 Experimental Results 42
4.8.1 Efficiency of Data Recording 42
4.8.2 Efy of Model Building and Prediction 43
4.9 Conclusion 44
5 Job Scheduling for Local SLA Management 47
5.1 Introduction 47
5.2 Basic Problem Statement 48
ii5.2.1 Problem Parameterization 48
5.2.2 Probabilistic Run Time 48
5.2.3 Scheduling Deadline-Constrained Jobs 50
5.3 More Complicated Scenarios 50
5.3.1 SLA Acknowledgement and Asynchronous Parameter Submission 50
5.3.2 User-specified Earliest Start Time 51
5.3.3 Lazy Termination 52
5.3.4 Alternative SLA Offers 53
5.4 Scheduling Phases 53
5.5 Scheduler Design and Implementation 54
5.6 Experimental Results 54
5.7 Related Work 58
5.7.1 Scheduling Algorithms 58
5.7.2 Performance Prediction Assisted Scheduling 58
5.8 Conclusion 58
6 Setting Up a Global Infrastructure 61
6.1 Overview 61
6.2 The Global Infrastructure 62
6.2.1 Establish SLA with Index and Broker Services 62
6.2.2 Enforce SLA with Gateway Services 63
6.3 Index S