Towards negotiable SLA-based QoS Support for Biomedical Data Services Gerhard Engelbrecht∗ , Jesus Bisbal∗ , Siegfried Benkner† and Alejandro F. Frangi∗ ∗
Center for Computational Imaging and Simulation Technologies in Biomedicine (CISTIB), Universitat Pompeu Fabra (UPF), Networking Biomedical Research Center on Bioengineering, Biomaterials and Nanomedicine (CIBER-BBN), and Instituci´o Catalana de Recerca i Estudis Avanc¸ats (ICREA) - c/ T´anger, 122-140, E08018 Barcelona, Spain Email:
[email protected] † University of Vienna, Department of Scientific Computing Nordbergstrasse 15, 3C, A-1090 Wien, Austria
To verify our assumption, we propose to advance wellknown QoS mechanisms from the computing domain [4] to be applied in the context of data access. At the core of this approach a QoS agreement is established between a data consumer and a data provider in advance to the actual data retrieval. Using such a QoS agreement meets the scientist’s objective to perform data access more efficiently by obtaining exactly the data they require, rather than querying a lot of data sources in a ”best-effort” fashion. Distributed e-Science infrastructures often adopt a service-oriented architecture (SOA), which utilizes dedicated data services such as presented in [2] to enable access to miscellaneous data. Data services are Web Services, which expose a data source such as a database, a set of files or another information system as a virtual data source to its clients. In SOA-based environments QoS is typically expressed with service level agreements (SLAs) in order to standardize the terms and conditions of a service consumption on a case-by-case basis. Accordingly, SLAs for data services comprise data-related guarantees, which are subject to negotiation and agreed prior to the actual data access. These guarantees are also referred to as service level objectives (SLOs), which in turn have to be customized for data access and retrieval. Our approach also introduces a set of specific SLOs, which can be negotiated between a client and a service in a request-offer fashion. Each service internally utilizes a set of estimation models to tackle the complexity of resolving requested SLOs against its actual capabilities of serving the request. Service composition and SLA aggregation are also considered in our approach by hierarchically combining data services extending the data mediation presented in [21]. In turn, this enables a data mediation service to consider only those data sources required to fulfill the requested SLOs. This QoS-based mediation capability is employed in the context of the @neurIST project [8] to demonstrate our approach and its advantages in a real world scenario. This paper is organized as follows: Related work is addressed in Section 2. Section 3 presents the basic principles
Abstract—Researchers in data intensive domains, like the Virtual Physiological Human initiative (VPH-I), are commonly overwhelmed with the vast and increasing amount of data available. Advanced studies in biomedicine and other domains often require a considerable amount of effort to achieve data access to a critical mass of relevant data to analyze the problem at hand. We aim to improve this situation and propose a novel application of Quality of Service (QoS) mechanisms for data services. This enables scientists to obtain exactly the data they require, rather than being spoilt for choice which data source might comprise suitable data. The proposed QoS support includes a negotiation model based on service level agreements (SLAs), which in turn comprises data-related service level objectives (SLOs) to express the required guarantees about the quantity or quality of data. Moreover a corresponding QoS management model is presented which resolves the complex process of the SLA generation within data access and data mediation services. The benefits of this approach are materialized in the context of the @neurIST data environment and an initial experimental evaluation demonstrates promising performance improvements in a real world scenario. Keywords-Quality of Service; SLA; SLO; Data Service; Data Quality;
I. I NTRODUCTION Technological advances in many scientific disciplines have led to the generation of vast amounts of potentially useful data to be analyzed and exploited. Scientists are typically spoilt for choice which data should actually be used [9]. Specifically, the Virtual Physiological Human (VPH) aims to conceptualize the human body as a single integrated system for the understanding of complex physiological processes [10]. The realisation of the VPH concept is, therefore, heavily dependent on the access, analysis, processing and modeling of heterogeneous data, which is usually distributed across many independent but inter-related data sources. Current practice tends to create customized databases for specific research domains and both the population and exploitation of these databases are conducted following a ”best-effort” policy [6]. In this paper, we hypothesize that the ability to carry out data-intensive research can be improved significantly by utilizing mechanisms to support Quality of Service (QoS).
978-1-4244-9349-4/10/$26.00 © 2010 IEEE
259
11th IEEE/ACM International Conference on Grid Computing
of the QoS support for data services and Section 4 demonstrates the QoS support in the context of the @neurIST project including an experimental evaluation and results. Section 5 concludes the paper and Section 6 outlines future work.
support invocation following a QoS negotiation protocol. Opposed to our work, these Web services do not represent compute or data-intensive applications or large-scale data resources. A primer infrastructure with respect to large data processing in Europe is coordinated by the European Grid Initiative (EGI)2 following up the EGEE projects [15]. EGI primarily deals with large-scale scientific data sets from scientific experiments represented as data files; access to databases and data integration are not considered. The requirements for handling large-scale volumes of data, access to individual records which have fields distributed over several, heterogeneous databases and have complex security demands are more evident in other emerging efforts like the Virtual Physiological Human (VPH) [10]. More specifically, our approach is demonstrated in the data infrastructure of the @neurIST project one of the early exponents of the VPH concept, which developed a complex information storage, access and processing infrastructure [5] .
II. R ELATED W ORK Quality of Service in the context of service oriented architectures (SOA) has received considerable scientific attention (e.g. recently by an EC-funded integrated project SLA@SOI1 ). While existing work on QoS has focused mainly on availability, reliability, security and cost, QoSissues related to data access and retrieval are rather less investigated. The work presented in [6] combines QoS and data access using SOA-principles. The work emphasizes QoS support for distributed query processing, suggesting QoS management at every stage of a query execution including planning, optimization and execution. As a consequence, significantly better performance results are achieved, as compared to systems based on a ”best-effort” policy. The work presented in [7] proposes a reputation-based system to predict one SLO for the response time of a query-execution in a Gridenvironment. Similarly, the work in [19] investigates the response time of query processing in parallel database management systems using an analytical approach. Our model supports an arbitrary set of SLOs and although we focus on SLOs addressing data quantity and quality, our approach can be combined with the above mentioned mechanisms for estimating the response time or cost. Work related to aggregation of SLOs is usually customized in a particular context. The work in [14] presents a composition algorithm for Web Services and SLAs while [20] proposes a formal model in the context of business processes to aggregate SLOs. Our approach also takes into account the requirements for the SLO aggregation specific to data retrieval and addresses the aggregation in the context of data mediation. Standardization efforts for SLAs and Web Services have lead to two main specifications: the Web Service Agreement (WS-Agreement) [1] promoted by the Open Grid forum (OGF) and the Web Service Level Agreement language and framework (WSLA) [13]. Both specifications are broadly adopted in many research projects and service-oriented infrastructures. Also both languages support the definition of arbitrary SLOs, which is required by our approach to express data-related SLOs. Although both WS-Agreement and WSLA may be used interchangeably, we initially adopted the latter driven by previous experience. An introductory work of QoS in Grid computing [16] addresses a generic QoS-based Web services architecture. The Web services are built upon QoS-aware components, which 1 SLA@SOI
III. Q O S SUPPORT MODEL The overall QoS support model comprises a client interacting with one or more QoS-aware services to establish an agreement on certain qualities of a service. Existing standards such as the WSLA and the WS-Agreement define this interaction as QoS support establishment phase, which takes place in advance to the actual service usage. Our overall QoS support model as described next, has been originally developed in the context of time-critical computing services [3] and is now extended for supporting data services.
Figure 1. Overview of the QoS support: QoS negotiation between a client and multiple services and QoS management in each service.
Figure 1 illustrates the QoS support model, distinguishing QoS negotiation and QoS management. On the left hand side the QoS negotiation is shown: a client negotiates with one or more services by exchanging requests and offers; and eventually, establishes a QoS agreement with one service provider. The right-hand side of the figure depicts the QoS management where each service provider individually attempts to satisfy a client’s request and to supply a corresponding QoS offer. A. Data-related service level objectives SLOs are specific measurable characteristics of a service, which are defined by an identifier (name) and a corresponding value, which expresses the actual objective. Typical 2 EGI:
project: http://sla-at-soi.eu/
260
http://www.egi.eu/
examples are availability maintaining certain threshold (e.g. 99 percent) or response time of less than some milliseconds. In case of data services the SLOs express measurable objectives or guarantees about the quantity or quality of the data being queried or the cost of answering the query, etc. The actual SLOs are specified in an SLA following the WSLA definition [13]. Although the QoS model is generic and consequently supports the definition of arbitrary SLOs, our approach is customized for a set of specific SLOs derived from a recent biomedical project [8]. The SLOs considered in our approach are outlined as follows: •
•
•
•
•
•
•
study, statistical significance is achieved when the number of subjects is in the order of several hundreds or thousands of entries. To achieve the specified number of subjects an SLO for cardinality of the returned data has to be defined. 2) Investigated subjects: This use-case emphasizes the cardinality of data a certain enquiry is performed on, contrarily to the first use-case, which focuses on the cardinality of the returned data. In order to perform an enquiry on a representative amount of reliable data, an SLO for the cardinality of the reliable data has to be defined. This challenge occurs frequently with association studies that examine if and how often two medical conditions occur together. Potential findings can only be verified with a representative cardinality of the underlying available and/or reliable data. Both use-cases might also be combined, e.g. if data of a minimum number of patients fulfilling a certain condition (cardinality of results) and the underlying reliable data the query is performed on (cardinality of reliable data) should not exceed a certain number of subjects. In this case, the relative frequency of e.g. a certain condition can be verified.
Cardinality of results: Quantity of the data returned for a particular query or rows returned in the SQLquery-response. Cardinality of reliable data: Quantity of reliable data the query should be performed on. Databases commonly include incomplete records of data (e.g. some attributes have no value) and only complete records are useful to be queried. The actual level of incompleteness is not considered, as biomedical researchers typically discard incomplete records at all, no matter how many entries are missing. Cardinality of inquired data: Quantity of total data the query should be performed on, no matter whether the records comprise all attributes or not. Response time: Actual time of delivering the results, which depends on how long it takes to execute the query. Cost: Price of a particular query execution, which usually follows a certain pricing model (e.g. constant cost per query or dynamic cost based on the quantity of the returned data). Diversity: The diversity of the used data sources specifies the number of different data sources exposed by a certain data service. This is usually provided by a data mediation service to guarantee that the mediated data originates from different sites or sources to maintain a certain level of diversity. Locality: Due to legal or administrative issues the queried data sources may be restricted to a particular country or organizational domain.
B. QoS negotiation The QoS negotiation can be outlined as a client’s attempt to establish a QoS agreement with a service provider. The client drives the process based on a request-offer model with a selected set of candidate services, which are obtained e.g. through a registry service. The basic scenario is shown in Figure 2.
Figure 2. Client-driven QoS negotiation with multiple services in a requestoffer fashion and a final confirmation of a specific offer.
The details of the QoS negotiation process can be outlined as follows: initially the client generates a concrete data request (e.g. a SQL-statement) and a QoS request following the WSLA specification with the requested objectives (SLOs). For example, a client specifies a specific SQLquery along with certain objectives such as the minimum number of SQL-result-rows and a maximum price. The data request and the objectives are then passed to each service, which internally utilizes individually the QoS management, explained in SectionIII-C, to achieve the clients constraints and returns an appropriate QoS offer. In domain-specific infrastructures, e.g. e-Health, typically multiple data providers exist, which expose certain data, such as patient data in a standardized format. In such
The most obvious SLOs about response time and price cover the typical business use cases to specify the actual cost of a data retrieval and the time when the data has to be available. The other SLOs are specifically related to data and basically target two general use-cases, which are particularly relevant in the context of clinical and VPH-related research: 1) Returned subjects: A clinical study typically requires data from a minimum amount of subjects to ensure statistical significance. For example, patient-data of a minimum number of subjects should be analyzed for a certain medical condition. The respective number of subjects from the technical point of view is arbitrary, but for a clinical research
261
an environment, clients negotiate with multiple services. Consequently, each QoS offer has a short expiration time and finally, it is up to the client to confirm a specific offer before it expires. If a QoS offer is confirmed by the client, a QoS agreement is established. Alternatively, the client may not confirm any of the offers and restart the negotiation with different objectives. In this context, the policy applied by the client may follow a certain strategy with iterative requests and offers in an auction-like fashion as advocated in [17] for computing services.
solved. This brute-force approach clearly does not scale for an increasing number of models. Still assuming that all dependencies can be resolved a directed acyclic graph with the models as nodes and the input/output objectives as edges can be created. Consecutively, feasible sequences to assess the models can be determined by using a topology sort, which can also be applied with an increasing number of models due to its linear runtime. However, interdependencies or cyclic dependencies may occur, which is typically the case with conflicting objectives. For example, the price is dependent on the cardinality of the results, while the cardinality of the results depends on the amount of money the data consumer quotes. In this case, the entire sequence or graph of models may be iteratively invoked according to a certain solving strategy. In turn, this leads to the second challenge.
C. QoS management Data services with QoS support rely on a complex mechanism to generate an offer in response to the client’s request. This offer generation process is performed by the QoS management. Internally, the requested SLOs are processed and the actual capabilities of the service with respect to each SLO are determined by using appropriate estimation models. Our approach does not prescribe the actual nature of an estimation model, but assumes that such a model is capable to assess one or more SLOs (output) and optionally requires other SLOs as input. In general, a combination of these estimation models enables the QoS management to create an appropriate offer. A formal description of the QoS management is defined as follows: S = {s1 . . . sk }
M = {m1 . . . mn }
Challenge #2: Conflicting objectives In this case, the QoS management has to perform a multicriteria optimization before supplying an offer. The complexity of the problem space indicates that there is no general scalable solution available, but several techniques can be used to find approximate answers. This include heuristics, genetic algorithms, mixed integer programming and linear programming (MIP/LP) or answer set programming (ASP). Our approach tackles both challenges, but initially attempts to determine an orchestration of models and to generate an offer to the client at all. Optimizations and advancements in both areas are subject to further investigation in the future.
St+1 = M ◦ St
S denotes the set of k individual SLOs si , 1 ≤ i ≤ k. Similarly, the estimation models can be described by the set M with a total number of n individual models. One model estimates one or more objectives (outputs), thus, the number of models is less or equal to the number of SLOs n ≤ k. On the other hand, a model may also depend on one or more SLOs (inputs), e.g. an SLO for the cost may only be determined knowing the cardinality of the results. These dependencies of the models and SLOs are formalized with progression of the models with the set of objectives. The models are iteratively assessed with the already available SLOs to determine not yet available SLOs. The actual order within this model progression is defined as orchestration of the models, which constitutes the first challenge in this context.
Figure 3. QoS management with a set of interdependent estimation models.
Figure 3 provides an overview of QoS management, and a sample orchestration of estimation models. In this particular case the models are interdependent, e.g. model m2 requires the output of model m1 and m4 as input and its output is required by model m3 . In such a setup, the models are assessed iteratively. As mentioned earlier according mechanisms are used to eventually determine the overall set of objectives the service is able to guarantee.
Challenge #1: Orchestration of the estimation models The orchestration of estimation models considers the required inputs and produced outputs of each model, which are typically dependent on other models. Assuming an orchestration of models can be found, the eventual output of the QoS management will be a full set of objectives the service is able to guarantee to the client. An initial approach to explore all possible permutations of the set of models (k! possibilities) ensures that a solution is found, if the system of dependencies can, indeed, be
IV. U SE CASE : @ NEUR IST Data-related QoS requirements can be motivated in any data-intensive domain. Specifically, in clinical studies, and in general, in advanced biomedical studies, conclusions can only be drawn if these are based on a sufficiently large number of subjects (e.g. patients). This scenario was indeed encountered in the @neurIST project.
262
management component to become QoS-aware. The QoS management comprises a set of estimation models according to the supported SLOs. These models are used to assess the SLOs and eventually provide an offer to the client. While the QoS management in a DAS basically assesses its estimation models to provide an offer, the DMS has to aggregate an offer based on the capacity of its associated DASes. Therefore, the DMS forwards the SLA request from the client to each associated DAS individually. Based on the received offers from the DASes, the DMS composes an offer that is returned to the client. The offer-composition of the DMS requires additional information for each SLO to perform a proper aggregation. For example, in the case of an SLO on the cardinality of result data, the DMS will need to add up the offers made by each DAS until the request can be satisfied. In this case, the information associated P with this SLO comprises an aggregation based on the function and the satisfaction condition ≥. Table I lists the additional information for each SLOs specified in Section III-A. Each objective is linked to an estimation function, which corresponds to an estimation model introduced in Section III-C, a satisfaction condition, which is usually a comparator derived from the associated predicate in the SLA, and an according aggregation function.
A. The @neurIST project @neurIST was an Information Society Technologies (IST) Integrated Project funded within the European Commissions Sixth Framework Programme. One of the major outcomes of the project was a service-oriented ICT infrastructure to improve the management and processing of data associated with the diagnosis and treatment of cerebral aneurysms. Although @neurIST focused on a particular disease, the infrastructure has been designed in a generic way to enable transparent access to and integration of heterogeneous data from various sources [5]. In the @neurIST project, the Clinical Reference Information Model (CRIM) [11] has been introduced as a uniform structure for all clinical patient data. Consequently, each clinical data provider in @neurIST - typically a hospital - exposed patient data according to the CRIM structure. Following the SOA-approach, the actual data access was accomplished through data access services (DAS). The final production system comprised a total of six different data access services exposing clinical patient data across Europe. The individual sites used different approaches to transform the data from their local hospital information systems (HIS) to the CRIM representation. Also a considerable effort was required by the clinical data providers to integrate heterogeneous patient data and to make it available via data access services in the CRIM structure. Finally, on top of these individual data access services exposing patient data, a data mediation service (DMS) has been created. The DMS federates the data from the different clinical sites into a large virtual database [5].
Table I SLA AGGREGATION INFORMATION Objective Cardinality of total/reliable/result data Cost of enquiry Response time of enquiry Diversity of data sources Locality of data sources
Estimation Satisfaction function Condition card cost resp dive loca
≥ ≤ ≤ = =
Aggregation
P P card(DASi )
cost(DASi ) max P resp(DASi ) V dive(DASi ) loca(DASi )
The additional information to enable the aggregation of SLAs in the context of data access (i.e. aggregation information for each data-related SLOs) is subject to an integrated SLA management framework and goes beyond the scope of the WSLA specification. Therefore, the definition and formalization of SLO-specific aggregation information, e.g. by expressing these semantics in a suitable declarative language, remains a challenge. The SLA@SOI project and research therein [18] will elaborate this issue in the context of creating a comprehensive SLA management framework.
Figure 4. @neurIST CRIM-based data access and mediation services. The abbreviation indicate the service type (DAS/DMS), the exposed data (CRIM) and the service provider.
Figure 4 depicts the structure of the @neurIST data access and mediation services for clinical patient data following the CRIM structure. Both, data access and mediation services, have the same interface and expose the same data structure (i.e. according to the CRIM). The data mediation service exposes virtually all clinical patient data available in the project. From the user’s perspective the data mediation services appears as a single data access point, but actually the data is retrieved from different sources.
C. Experimental Evaluation An initial evaluation of a preliminary implementation of our approach showed that the performance of the DMS with respect to processing of a query can be improved significantly incorporating QoS support. An experiment has been conducted with 20 sample queries that have been selected and executed against the DMS in a best-effort fashion (i.e. no QoS) and with QoS support.
B. QoS in @neurIST data services Realizing the QoS support, as described in Section III-C, in the context of the @neurIST infrastructure requires to extend all data services (DAS and DMS) with a QoS
263
Figure 5.
Figure 6.
Query processing time of 20 sample queries using a DMS.
Performance improvement of a DMS with QoS support.
being in the range of megabytes. This situation can be explained by the overhead introduced by the QoS negotiation and QoS management, which usually ranges from 5% to 10% of the total query processing time. On other hand, this also indicates that improving the QoS negotiation and QoS management remains challenging. The overall trend of this experiment shows, that for small-sized results, introducing QoS-support significantly improves the performance of the query processing. In general this can be explained by the fact that the DMS may not need to accumulate data from all associated DASes contrarily to the best effort solution. The request could be satisfied with only a subset of the associated DASes. The presented experiment focuses on improving the performance, which, indeed, is only a single feature of the overall QoS support. However, the evaluation also shows that the QoS support can be applied in a real world scenario and that scientists can benefit beyond performance by retrieving only relevant and useful data for their research.
The results are shown in Figures 5 and 6. For this experiment we selected basic SQL queries (i.e. SELECT * FROM tablename) to retrieve all available data from a particular table defined in the CRIM. Although this setup appears simplistic, the clinical and research enquiries observed frequently in the project, usually combine different tables and add simple conditions (i.e. WHERE-clauses). These queries are numbered from 1 to 20 and depicted on the horizontal axis of the Figures 5 and 6. Moreover, the queries are sorted increasingly according to the size of their results, which range from a few kilobytes (query #1) to a few megabytes (query #20). We intentionally put no explicit scale for the size of the results, as different volumes of data is being generated with and without QoS support. For the QoS-enabled query processing an SLO for the total number of results has been specified with 50 and 100 subjects, which are depicted in the figures with QoS/50 and QoS/100, respectively. While the horizontal axis poses the actual query, the vertical axis shows the processing time in seconds using the best-effort and QoS-enabled option in Figures 5 as well as the performance improvement against the best-effort option in Figures 6. The obtained results indicate that applying QoS support, even in this rather simple scenario, improves the overall performance of the query processing performed by the data mediation service significantly. The quantification of the resulting data with a total of 50 or 100 subjects is fairly usual in browser-style applications in order to reduce the number of results to display (c.f. a Google-query only displays 10 results by default). Compared to the best-effort data retrieval the QoS/50 data retrieval improves the query processing time up to over 60% for small-sized results and between 10% and 30% for larger results. Due to larger data to process, obviously 100 results are larger than 50, the QoS/100 data retrieval requires more processing time and even performs worse compared to the best-effort data retrieval for results
V. C ONCLUSIONS We have presented a negotiable QoS support for data services to guarantee data-related service level objectives (SLOs) such as cardinality of the results in advance to the actual data retrieval. In the presented application scenario of @neurIST, researchers are supported while performing clinical studies to gain access to exactly the critical mass of data that is required for their problem at hand. Consequently, clinical and research studies in data-intensive domains can be carried out with less effort to accomplish the required data access. With respect to the internals of QoS-aware data service the complex process of generating an offer with the QoS management has been investigated with a single service and in the context of aggregating multiple services. The presented QoS support has been materialized in the @neurIST data
264
infrastructure to demonstrate our approach in a real world scenario. Also in this context, an initial experiment showed that the performance of query processing can be improved significantly.
[7] de Carvalho Costa, R. L. and Furtado, P. QoS-Oriented Reputation-Aware Query Scheduling in Data Grids. In EuroPar 2008, pages 489-498, Springer-Verlag, Berlin, Heidelberg, 2008.
VI. F UTURE WORK
[8] Frangi, A.F., and Ruiz, A. Understanding Cerebral Aneurysms: The @neurIST Project. In Special Issue on the Digital Patient, ERCIM Newsletter n. 69: 16-17, April 2007.
For the future we plan to elaborate the existing system as well as explore new directions the system can be applied. Advancing the presented system includes addressing the identified challenges towards more sophisticated negotiation mechanisms, different QoS management optimization strategies, and enhanced dynamic data mediation. Moreover, we will investigate a formal definition of the SLA aggregation and link semantic information to data-related SLOs. Concerning new directions, we plan to apply our approach in combination with virtualization techniques in Cloud computing. Virtualizing data as a service utilizing appropriate Cloud environments also constitutes a major challenge and Quality of Service remains a hot topic in this context.
[9] Hey, T. and Trefethen, A. The data deluge: an e-Science perspective. In Grid computing: making the global infrastructure a reality (F. Berman, G. Fox and T. Hey eds.), pages 809-824, Wiley and Sons, 2003. [10] Hunter, P.J. et al. A vision and strategy for the virtual physiological human in 2010 and beyond. In Philosophical Transactions of the Royial Society A 368 (1920): 2595-2614, 2010. [11] Iavindrasana, J. at al. Design of a Decentralized Reusable Research Database Architecture to Support Data Acquisition in Large Research Projects. In Studies in health technology and informatics 129 (Pt 1): 325-329, 2007.
ACKNOWLEDGMENT
[12] Jin, H. and Wu H. Semantic-enabled Specification for Web Services Agreement. In Web Services Practices 1 (1): 13-20, 2005.
This work was partially funded by the Integrated Project @neurIST (FP6-IST-027703), which is cofinanced by the European Commission, and by the Spanish Ministry of Innovation and Science (MICINN) through the cvREMOD project (CEN-20091044) under the CENIT programme of the Industrial and Technological Development Center.
[13] Keller, A. and Ludwig, H.. The WSLA Framework: Specifying and Monitoring Service Level Agreements for Web Services. In Network and Systems Management 11 (1): 57-81, 2003. [14] Ko, J.M., Kim, C.O. and Kwon, I.H. Quality-of-service oriented web service composition algorithm and planning architecture In Systems and Software, 81 (11): 2079-2090, 2008.
R EFERENCES [1] Andrieux, A. et al. Web Service Agreement Specification (WS-Agreement). Open Grid Forum specification: http://www.ogf.org/documents/GFD.107.pdf, 2007.
[15] Laure, E. and Jones, B. Enabling Grids for e-Science: The EGEE Project. In Grid Computing: Infrastructure, Service, and Application, CRC Press, 2009.
[2] Antonioletti, M., Atkinson, M. et al. The design and implementation of Grid database services in OGSA-DAI: Research Articles. In Concurrency and Computing: Practice & Experience 17 (2-4): 357-376, 2005.
[16] Menasce, D.A. and Casalicchio, E. QoS in Grid Computing. In IEEE Internet Computing 8 (4): 95-87, 2004. [17] Middleton, S. E., Surridge, M., Benkner, S. and Engelbrecht, G. Quality of service negotiation for commercial medical Grid services, In Grid Computing, 5 (4): 429-447, 2007.
[3] Benkner, S., Engelbrecht, G., Middleton, S. E., Brandic, I. and Schmidt, R. End-to-End QoS Support for a Medical Grid Service Infrastructure. In New Generation Computing, Computing Paradigms and Computational Intelligence, Special Issue on Life Science Grid Computing 25 (4): 355-372, 2007.
[18] Theilmann, W. and Baresi, L. Multi-level SLAs for Harmonized Management in the Future Internet. In Towards the Future Internet - A European Research Perspective, pages 193202, IOS Press, 2009.
[4] Benkner, S., Engelbrecht, G., K¨ohler, M. and W¨ohrer, A. Virtualizing Scientific Applications and Data Sources as Grid Services. In Cyberinfrastructure Technologies and Applications (J. Cao eds.): pages 81-111, Nova Science Publishers, New York, USA, 2008.
[19] Tomov, N., Dempster, E., Williams, M.H., Burger, A., Taylor, H., King, P.J.B. and Broughton, P. Analytical response time estimation in parallel relational database systems. In Parallel Computing 30 (2): 249-283, 2004.
[5] Benkner, S., Arbona, A. et al. @neurIST - Infrastructure for Advanced Disease Management Through Integration of Heterogeneous Data, Computing, and Complex Processing Services. In IEEE Transactions on Information Technology in Biomedicine in press, 2010.
[20] Unger, T., Leymann, F., Mauchart, S. and Scheibler, T. Aggregation of Service Level Agreements in the Context of Business Processes. In Enterprise Distributed Object Computing Conference 2008: 43-52, Los Alamitos, CA, USA, 2008.
[6] R. Braumandl and A. Kemper and D. Kossmann. Quality of Serivce in an Information Economy. In ACM Transactions on Internet Technology 3 (4): 291-333, 2003.
[21] W¨ohrer, A., Brezany, P. and Tjoa A. Novel mediator architectures for Grid information systems. In Future Generation Computer Systems, 21 (1): 107-114, 2005.
265 View publication stats
