Designing High Performance Web-Based Computing Services to Promote Telemedicine Database Management System

Transcription

1 Designing High Perforance Web-Based Coputing Services to Proote Teleedicine Database Manageent Syste Isail Hababeh 1, Issa Khalil 2, and Abdallah Khreishah 3 1: Coputer Engineering & Inforation Technology, Geran-Jordanian University, Jordan, Isail.Hababeh@gju.edu.jo 2: Qatar Coputing Research Institute (QCRI), Qatar Foundation, Doha, Qatar, ikhalil@qf.org.qa 3: Departent of Electrical & Coputer Engineering, New Jersey Institute of Technology, USA, Abdallah@njit.edu Corresponding Author: Issa Khalil, Eail: ikhalil@qf.org.qa Abstract Many web coputing systes are running real tie database services where their inforation change continuously and expand increentally. In this context, web data services have a ajor role and draw significant iproveents in onitoring and controlling the inforation truthfulness and data propagation. Currently, web teleedicine database services are of central iportance to distributed systes. However, the increasing coplexity and the rapid growth of the real world healthcare challenging applications ake it hard to induce the database adinistrative staff. In this paper, we build an integrated web data services that satisfy fast response tie for large scale Tele-health database anageent systes. Our focus will be on database anageent with application scenarios in dynaic teleedicine systes to increase care adissions and decrease care difficulties such as distance, travel, and tie liitations. We propose threefold approach based on data fragentation, database web sites clustering and intelligent data distribution. This approach reduces the aount of data igrated between web sites during applications execution; achieves cost-effective counications during applications processing and iproves applications response tie and throughput. The proposed approach is validated internally by easuring the ipact of using our coputing services techniques on various perforance features like counications cost, response tie, and throughput. The external validation is achieved by coparing the perforance of our approach to that of other techniques in the literature. The results show that our integrated approach significantly iproves the perforance of web database systes and outperfors its counterparts. Keywords: Web Teleedicine Database Systes (WTDS); database fragentation; data distribution; sites clustering. 1

2 1. Introduction The rapid growth and continuous change of the real world software applications have provoked researchers to propose several coputing services techniques to achieve ore efficient and effective anageent of web teleedicine database systes (WTDS). Significant research progress has been ade in the past few years to iprove WTDS perforance. In particular, databases as a critical coponent of these systes have attracted any researchers. The web plays an iportant role in enabling healthcare services like teleedicine to serve inaccessible areas where there are few edical resources. It offers an easy and global access to patients data without having to interact with the in person and it provides fast channels to consult specialists in eergency situations. Different kinds of patient s inforation such as ECG, teperature, and heart rate need to be accessed by eans of various client devices in heterogeneous counications environents. WTDS enable high quality continuous delivery of patient s inforation wherever and whenever needed. Several benefits can be achieved by using web teleedicine services including: edical consultation delivery, transportation cost savings, data storage savings, and obile applications support that overcoe obstacles related to the perforance (e.g. bandwidth, battery life, and storage), security (e.g. privacy, and reliability), and environent (e.g. scalability, heterogeneity, and availability). The objectives of such services are to: (i) develop large applications that scale as the scope and workload increases, (ii) achieve precise control and onitoring on edical data to generate high teleedicine database syste perforance, (iii) provide large data archive of edical data records, accurate decision support systes, and trusted event-based notifications in typical clinical centers. Recently, any researchers have focused on designing web edical database anageent systes that satisfy certain perforance levels. Such perforance is evaluated by easuring the aount of relevant and irrelevant data accessed and the aount of transferred edical data during transactions processing tie. Several techniques have been proposed in order to iprove teleedicine database perforance, optiize edical data distribution, and control edical data proliferation. These techniques believed that high perforance for such systes can be achieved by iproving at least one of the database web anageent services, naelydatabase fragentation, data distribution, web sites clustering, distributed caching, and database scalability. However, the intractable tie coplexity of processing large nuber of edical transactions and anaging huge nuber of counications ake the design of such ethods a non-trivial task. Moreover, none of the existing ethods consider the three-fold services together which akes the ipracticable in the field of web database systes. Additionally, using ultiple edical services fro different web database providers ay 2

3 not fit the needs for iproving the teleedicine database syste perforance. Furtherore, the services fro different web database providers ay not be copatible or in soe cases it ay increase the processing tie because of the constraints on the network [1]. Finally, there has been lack in the tools that support the design, analysis and cost-effective deployents of web teleedicine database systes [1]. Designing and developing fast, efficient, and reliable incorporated techniques that can handle huge nuber of edical transactions on large nuber of web healthcare sites in near optial polynoial tie are key challenges in the area of WTDS. Data fragentation, web sites clustering, and data allocation are the ain coponents of the WTDS that continue to create great research challenges as their current best near optial solutions are all NP-Coplete. To iprove the perforance of edical distributed database systes, we incorporate data fragentation, web sites clustering, and data distribution coputing services together in a new web teleedicine database syste approach. This new approach intends to decrease data counication, increase syste throughput, reliability, and data availability. The decoposition of web teleedicine database relations into disjoint fragents allows database transactions to be executed concurrently and hence iniizes the total response tie. Fragentation typically increases the level of concurrency and, therefore, the syste throughput. The benefits of generating teleedicine disjoint fragents cannot be deeed unless distributing these fragents over the web sites, so that they reduce counication cost of database transactions. Database disjoint fragents are initially distributed over logical clusters (a group of web sites that satisfy a certain physical property, e.g. counications cost). Distributing database disjoint fragents to clusters where a benefit allocation is achieved, rather than allocating the fragents to all web sites, have an iportant ipact on database syste throughput. This type of distribution reduces the nuber of counications required for query processing in ters of retrieval and update transactions; it has always a significant ipact on the web teleedicine database syste perforance. Moreover, distributing disjoint fragents aong the web sites where it is needed ost, iproves database syste perforance by iniizing the data transferred and accessed during the execution tie, reducing the storage overheads, and increasing availability and reliability as ultiple copies of the sae data are allocated. Database partitioning techniques ai at iproving database systes throughput by reducing the aount of irrelevant data packets (fragents) to be accessed and transferred aong different web sites. However, data fragentation raises soe difficulties; particularly when web teleedicine database applications have contradictory requireents that avert breakdown of the relation into utually exclusive fragents. Those applications whose views are defined on ore than 3

4 one fragent ay suffer perforance ruin. In this case, it ight be necessary to retrieve data fro two or ore fragents and take their join, which is costly [31]. Data fragentation technique describes how each fragent is derived fro the database global relations. Three ain classes of data fragentation have been discussed in the literature; horizontal [2][3], vertical [4][5], and hybrid [6][7]. Although there are various schees describing data partitioning, few are known for the efficiency of their algoriths and the validity of their results [33]. The Clustering technique identifies groups of network sites in large web database systes and discovers better data distributions aong the. This technique is considered to be an efficient ethod that has a ajor role in reducing the aount of transferred and accessed data during processing database transactions. Accordingly, clustering techniques help in eliinating the extra counications costs between web sites and thus enhances distributed database systes perforance [32]. However, the assuptions on the web counications and the restrictions on the nuber of network sites, ake clustering solutions ipractical [16][31]. Moreover, soe constraints about network connectivity and transactions processing tie bound the applicability of the proposed solutions to sall nuber of clusters [9][10]. Data distribution describes the way of allocating the disjoint fragents aong the web clusters and their respective sites of the database syste. This process addresses the assignent of each data fragent to the distributed database web sites [8][17][18][21]. Data distribution related techniques ai at iproving distributed database systes perforance. This can be accoplished by reducing the nuber of database fragents that are transferred and accessed during the execution tie. Additionally, Data distribution techniques attept to increase data availability, elevate database reliability, and reduce storage overhead [11][27]. However, the restrictions on database retrieval and update frequencies in soe data allocation ethods ay negatively affect the fragents distribution over the web sites [20]. In this work, we address the previous drawbacks and propose a three-fold approach that anages the coputing web services that are required to proote teleedicine database syste perforance. The ain contributions are: Develop a fragentation coputing service technique by splitting teleedicine database relations into sall disjoint fragents. This technique generates the iniu nuber of disjoint fragents that would be allocated to the web servers in the data distribution phase. This in turn reduces the data transferred and accessed through different web sites and accordingly reduces the counications cost. 4

5 Introduce a high speed clustering service technique that groups the web teleedicine database sites into sets of clusters according to their counications cost. This helps in grouping the web sites that are ore suitable to be in one cluster to iniize data allocation operations, which in turn helps to avoid allocating redundant data. Propose a new coputing service technique for teleedicine data allocation and redistribution services based on transactions processing cost functions. These functions guarantee the iniu counications cost aong web sites and hence accoplish better data distribution copared to allocating data to all web sites evenly. Develop a user-friendly experiental tool to perfor services of teleedicine data fragentation, web sites clustering, and fragents allocation, as well as assist database adinistrators in easuring WTDS perforance. Integrate teleedicine database fragentation, web sites clustering, and data fragents allocation into one scenario to accoplish ultiate web teleedicine syste throughput in ters of concurrency, reliability, and data availability. We call this scenario Integrated-Fragentation-Clustering- Allocation (IFCA) approach. Figure 1 depicts the architecture of the proposed teleedicine IFCA approach. Web Teleedicine Database Syste Adinistrator WS 1 Phase 6: Executing Allocation Technique on Clusters & Allocate Fragents to Clusters WS 2 WS 3.. Phase 7: Executing Allocation Technique on Sites & Allocate Fragents to Sites Phase 5: Executing Clustering Technique & Generating Clusters Phase 4: Executing Fragentation Technique & Generating Disjoint Fragents DF1 DF2 DF3. Phase 1: Requesting Data for Processing Web Teleedicine Database Sites Figure 1: IFCA Coputing Services Architecture In Figure 1, the data request is initiated fro the teleedicine database syste sites. The requested data is defined as SQL queries that are executed on the database relations to generate data set records. Soe of these data records ay be overlapped or even redundant, which increase the I/O transactions processing tie and so the syste counications overhead. To solve this proble, we execute the proposed fragentation technique which generates teleedicine disjoint fragents that represent the iniu nuber of data records. The web teleedicine database sites are grouped into clusters by using our clustering service technique in a phase prior to data allocation. The purpose of this clustering is to reduce the counications cost needed for data allocation. Accordingly, the proposed allocation service DSR1 Phase 2: Defining Queries Phase 3: Executing Queries & Generating Data Set of Records QR3. QR2 QR1 DSR2 DSR3 DSR4 5

6 technique is applied to allocate the generated disjoint fragents at the clusters that show positive benefit allocation. Then the fragents are allocated to the sites within the selected clusters. Database adinistrator is responsible for recovering any site failure in the WTDS. The reainder of the paper is organized as follows. Section 22 suarizes the related work. Basic concepts of the web teleedicine database settings and assuptions are discussed in Section 3. Teleedicine coputation services and estiation odel are discussed in Section 4. Experiental results and perforance evaluation are presented in Section 35. Finally, in Section 6, we draw conclusions and outline the future work. 2. Related Work Many research works have attepted to iprove the perforance of distributed database systes. These works have ostly investigated fragentation, allocation and soeties clustering probles. In this section, we present the ain contributions related to these probles, discuss and copare their contributions with our proposed solutions Data Fragentation With respect to fragentation, the unit of data distribution is a vital issue. A relation is not appropriate for distribution as application views are usually subsets of relations *31+. Therefore, the locality of applications accesses is defined on the derivative relations subsets. Hence it is iportant to divide the relation into saller data fragents and consider it for distribution over the network sites. The authors in [8] considered each record in each database relation as a disjoint fragent that is subject for allocation in a distributed database sites. However, large nuber of database fragents is generated in this ethod, causing a high counication cost for transitting and processing the fragents. In contrast to this approach, the authors in [11] considered the whole relation as a fragent, not all the records of the fragent have to be retrieved or updated, and a selectivity atrix that indicates the percentage of accessing a fragent by a transaction is proposed. However, this research suffers fro data redundancy and fragents overlapping Clustering Web Sites Clustering service technique identifies groups of networking sites and discovers interesting distributions aong large web database systes. This technique is considered as an efficient ethod that has a ajor role in reducing transferred and accessed data during transactions processing [9]. Moreover, grouping distributed network sites into clusters helps to 6

7 eliinate the extra counication costs between the sites and then enhances the distributed database syste perforance by iniizing the counication costs required for processing the transactions at run tie. In a web database syste environent where the nuber of sites has expanded treendously and aount of data has increased enorously, the sites are required to anage these data and should allow data transparency to the users of the database. Moreover, to have a reliable database syste, the transactions should be executed very fast in a flexible load balancing database environent. When the nuber of sites in a web database syste increases to a large scale, the proble of supporting high syste perforance with consistency and availability constraints becoes crucial. Different techniques could be developed for this purpose; one of the is web sites clustering. Grouping web sites into clusters reduces counications cost and then enhances the perforance of the web database syste. However, clustering network sites is still an open proble and the optial solution to this proble is NP- Coplete [12]. Moreover, in case of a coplex network where large nubers of sites are connected to each other, a huge nuber of counications are required, which increases the syste load and degrades its perforance. The authors in [13] have proposed a hierarchical clustering algorith that uses siilarity upper approxiation derived fro a tolerance(siilarity) relation and based on rough set theory that does not require any prior inforation about the data. The presented approach results in rough clusters in which an object is a eber of ore than one cluster. Rough clustering can help researchers to discover ultiple needs and interests in a session by looking at the ultiple clusters that a session belongs to. However, in order to carry out rough clustering, two additional requireents, naely, an ordered value set of each attribute and a distance easure for clustering need to be specified [14]. Clustering coefficients are needed in any approaches in order to quantify the structural network properties. In [15], the authors proposed higher order clustering coefficients defined as probabilities that deterine the shortest distance between any two nearest neighbors of a certain node when neglecting all paths crossing this node. The outcoes of this ethod declare that the average shortest distance in the node s neighborhood is saller than all network distances. However, independent constant values and natural logarith function are used in the shortest distance approxiation function to deterine the clustering echanis, which results in generating sall nuber of clusters Data Allocation (Distribution) Data allocation describes the way of distributing the database fragents aong the clusters and their respective sites in distributed database systes. This process addresses the assignent of network node(s) to each fragent [8]. However, 7

8 finding an optial data allocation is NP-coplete proble [12]. Distributing data fragents aong database web sites iproves database syste perforance by iniizing the data transferred and accessed during execution, reducing the storage overhead, and increasing availability and reliability where ultiple copies of the sae data are allocated. Many data allocation algoriths are described in the literature. The efficiency of these algoriths is easured in ter of response tie. Authors in [19] proposed an approach that handles the full replication of data allocation in database systes. In this approach, a database file is fully copied to all participating nodes through the aster node. This approach distributes the sequences through fragents with a round-robin strategy for sequence input set already ordered by size, where the nuber of sequences is about the sae and nuber of characters at each fragent is siilar. However, this replicated schea does not achieve any perforance gain when increasing the nuber of nodes. When a non-previously deterined nuber of input sequences are present, the replication odel ay not be the best solution and other fragentation strategies have to be considered. In [20], the author has addressed the fragent allocation proble in web database systes. He presented an integer prograing forulations for the non-redundant version of the fragent allocation proble. This forulation is extended to address probles, which have both storage and processing capacity constraints. In this ethod, the constraints essentially state that there has been exactly one copy of a fragent across all sites, which increase the risk of data inconsistency and unavailability in case of any site failure. However, the fragent size is not addressed while the storage capacity constraint is one of the ajor objectives of this approach. In addition, the retrieval and update frequencies are not considered in the forulations, they are assued to be the sae, which affects the fragents distribution over the sites. Moreover, this research is liited by the fact that none of the approaches presented have been ipleented and tested on a real web database syste. A dynaic ethod for data fragentation, allocation, and replication is proposed in [25]. The objective of this approach is to iniize the cost of access, re-fragentation, and reallocation. DYFRAM algorith of this ethod exaines accesses for each replica and evaluates possible re-fragentations and reallocations based on recent history. The algorith runs at given intervals, individually for each replica. However, data consistency and concurrency control are not considered in DYFRAM. Additionally, DYFRAM doesn t guarantee data availability and syste reliability when all sites have negative utility values. In [28], the authors present a horizontal fragentation technique that is capable of taking a fragentation decision at the initial stage, and then allocates the fragents aong the sites of DDBMS. A odified atrix MCRUD is constructed by placing predicates of attributes of a relation in rows and applications of the sites of a DDBMS in coluns. 8

9 Attribute locality precedence ALP; the value of iportance of an attribute with respect to sites of distributed database is generated as a table fro MCRUD. However, when all attributes have the sae locality precedence, the sae fragent has to be allocated in all sites, and a huge data redundancy occurs. Moreover, the initial values of frequencies and weights don t reflect the actual ones in real systes, and this ay affect the nuber of fragents and their allocation accordingly. The authors in [29] presented a ethod for odeling the distributed database fragentation by using UML 2.0 to iprove applications perforance. This ethod is based on a probability distribution function where the execution frequency of a transaction is estiated ainly by the ost likely tie. However, the ost likely tie is not deterined to distinguish the priorities between transactions. Furtherore, no perforance evaluations are perfored and no significant results are generated fro this ethod. A database tool shown in [30] addresses the proble of designing DDBs in the context of the relational data odel. Conceptual design, fragentation issues, as well as the allocation proble are considered based on other ethods in the literature. However, this tool doesn t consider the local optiization of fragent allocation proble over the distributed network sites. In addition, any design paraeters need to be estiated and entered by designers where different results ay be generated for the sae application case. Our fragentation approach circuvents the probles associated with the aforeentioned studies by introducing a coputing service technique that generates disjoint fragents, avoids data redundancy and considers that all records of the fragent are retrieved or updated by a transaction. By such a technique, less counication costs are required to transit and process the disjoint database fragents. In addition, by applying the clustering service technique, the coplexity of allocation proble is significantly reduced. Therefore, the intractable solution of fragent allocation is turned out into fragent distribution aong the clusters and then replicating it aong the related sites Coercial Outsource Databases This section investigates current coercial outsource databases and copares the with our IFCA. The coercial outsource databases support enorous data storage architecture that distributes storage and processing across several servers. It can be used to address web database syste perforance and scalability requireents. Aazon Dynao [34] is a cloud distributed database syste with high available and scalable capabilities. In contrast to traditional relational distributed database anageent systes, Dynao only supports database applications with siple read and write queries on data records. However, in Dynao availability is ore iportant than consistency. Accordingly, in certain ties, data consistency can t be guaranteed. MongoDB [35] is an open source, docuent oriented 9

10 cloud distributed database developed by the 10gen software copany. MongoDB is highly available and scalable syste that supports search by fields in the docuents. However, MongoDB uses a special query optiizer to ake queries ore efficient, and executes different query plans concurrently, which increases the search tie coplexity. Google BigTable [36] is a cloud distributed database syste built on Google file syste and a highly available and persistent distributed lock service. The BigTable data odel is designed as a distributed ultidiensional (row key, colun key, tiestaps) sorted ap, and ipleented on three parts: tablet servers, client library, and aster server. However, as the failure of the aster server results in the failure of the whole database syste, and thus another server usually backs up the aster server. This incurs additional costs to the cloud distributed database syste in addition to the costs of operating and aintaining the new servers. Due to their contrast in priorities, architecture, data consistency, and search tie coplexity copared to our IFCA, the previous outsource databases can t optiize and secure teleedicine data in ters of data fragentation, clustering web sites, and data distribution altogether as our approach successfully does. Table 1 copares our approach with outsource approaches in ters of integrity, reliability, counications overhead, anageability, data consistency, security, and perforance evaluation. Table 1. Coparison between existing ethods in the literature and the proposed approach Method # Integrity Reliability Counication overhead 8, 11 No, ajority doesn t apply clustering and allocation techniques 13,15 No, ajority doesn t apply fragentation and allocation techniques 19,20,25, No, ajority 28,29, 30 doesn t apply fragentation and clustering techniques 34, 35, No, ajority have 36 their own strategy for data allocation, but doesn t apply fragentation and clustering techniques IFCA Yes, full integrated fragentation, clustering, and allocations Poor due to fragents redundancy and Restrictions on the final # of fragents High because each site is considered in ultiple clusters Poor due to fault tolerant risk High due to the control of cloud services provider High since data available at ost benefit sites High traffic to ultiple sites, and soe delays are tolerated High due to site duplication in different clusters Low due to the sall # of data allocations Low as the cost is per use in a unit of tie Low due to place data where the least counication cost is hold Manageability Data consistency Security Perforance evaluation Difficult due to fragents overlapping and redundancy Not easy as the # of sites increase and to find the shortest distance algoriths Easy for each site, until there is a need to share data across sites Easy since it is supported by the cloud service provider Easy because the coplete data life cycle can be controlled by the provided techniques Low due to outcoes of the fragentation Inconsistencies are not tolerated Presents data consistency for each site, but not across sites High as it is the responsibility of the cloud service provider High due to the allocation technique in this approach that keep data consistency Secured Intractable coplexity barrier perforance are to Secured Poor due to redundant data in ultiple sites Secured Not secured since it is controlle d by cloud services provider Secured No Guarantees of high perforance as # of sites becoe intractable Acceptable level of perforance due to ore processing required for securing data High as the threefold techniques guarantees iniu data records allocated to sites with least co. costs 10

11 3. Teleedicine IFCA Assuptions and Definitions Incorporating database fragentation, web database sites clustering, and data fragents coputing services allocation techniques in one scenario distinguishes our approach fro other approaches. The functionality of such approach depends on the settings, assuptions, and definitions that identify the WTDS ipleentation environent, to guarantee its efficiency and continuity. Below are the description of the IFCA settings, assuptions, and definitions Web Architecture and Counications Assuptions The teleedicine IFCA approach is designed to support web database provider with coputing services that can be ipleented over ultiple servers, where the data storage, counication and processing transactions are fully controlled, costs of counication are syetric, and the patients inforation privacy and security are et. We propose fully connected sites on a web teleedicine heterogeneous network syste with different bandwidths; 128 kbps, 512 kbps, or ultiples. In this environent, soe servers are used to execute the teleedicine queries triggered fro different web database sites. Few servers are run the database progras and perfor the fragentation-clusteringallocation coputing services while the other servers are used to store the database fragents. Counications cost (s/byte) is the cost of loading and processing data fragents between any two sites in WTDS. To control and siplify the proposed web teleedicine counication syste, we assue that counication costs between sites are syetric and proportional to the distance between the. Counication costs within the sae site are neglected Fragentation and Clustering Assuptions Teleedicine queries are triggered fro web servers as transactions to deterine the specific inforation that should be extracted fro the database. Transactions include but not liited to: read, write, update, and delete. To control the process of database fragentation and to achieve data consistency in the teleedicine database syste, IFCA fragentation service technique partitions each database relation according to the Inclusion-Integration-Disjoint assuptions where the generated fragents ust contain all records in the database relations, the original relation should be able to be fored fro its fragents, and the fragents should be neither repeated nor intersected. The logical clustering decision is defined as a Logical value that specifies whether a web site is included or excluded fro a certain cluster, based on the counications cost range. The counications cost range is defined as a value (s/byte) that specifies how uch tie is allowed for the web sites to transit or receive their data to be considered in the sae cluster, this value is deterined by the teleedicine database adinistrator. 11

12 3.3. Fragents Allocation Assuptions The allocation decision value ADV is defined as a logical value (1, 0) that deterines the fragent allocation status for a specific cluster. The fragents that achieve allocation decision value of (1) are considered for allocation and replication process. The advantage that can be generated fro this assuption is that, ore counications costs are saved due to the fact that the fragents locations are in the sae place where it is processed, hence iprove the WTDS perforance. On the other hand, the fragents that carry out allocation decision value of (0) are considered for allocation process only in order to ensure data availability and fault-tolerant in the WTDS. In this case, each fragent should be allocated to at least one cluster and one site in this cluster. The allocation decision value ADV is assued to be coputed as the result of the coparison between the cost of allocating the fragent to the cluster and the cost of not allocating the fragent to the sae cluster. The allocation cost function is coposed of the following sub-cost functions that are required to perfor the fragent transactions locally: cost of local retrieval, cost of local update to aintain consistency aong all the fragents distributed over the web sites, and cost of storage, or cost of reote update and reote counications (for reote clusters that do not have the fragent and still need to perfor the required transactions on that fragent). The not allocation cost function consists of the following sub-cost functions: cost of local retrieval and cost of reote retrievals required to perfor the fragent transactions reotely when the fragent is not allocated to the cluster. 4. Teleedicine IFCA Coputation Services and Estiation Model In the following subsections, we present our IFCA and provide atheatical odels of its coputations services Fragentation Coputing Service To control the process of database fragentation and aintain data consistency, the fragentation technique partitions each database relation into data set records that guarantee data inclusion, integration and non-overlapping. In a WTDS, neither coplete relation nor attributes are suitable data units for distribution, especially when considering very large data. Therefore, it is appropriate to use data fragents that would be allocated to the WTDS sites. Data fragentation is based on the data records generated by executing the teleedicine SQL queries on the database relations. The fragentation process goes through two consecutive internal processes: (i) Overlapped and redundant data records fragentation and (ii) Non-overlapped data records fragentation. 12

13 The fragentation service generates disjoint fragents that represent the iniu nuber of data records to be distributed over the web sites by the data allocation service. The proposed fragentation Service architecture is described through Input-Processing-Output phases depicted in Figure 2. Based on this fragentation service, the global database is partitioned into disjoint fragents. The overlapped and redundant data records fragentation process is described in Table 2. In this algorith, database fragentation starts with any two rando data fragents (i,j) with intersection records and proceeds as follows: If intersection exists between the two fragents, three disjoint fragents will be generated: Fk= Fi Fj, F k+1 = Fi Fk, and F k+2 = Fj- F. The first contains the coon records, the second contains the records in the first fragent but not in the second fragent, and the third contains the records in the second fragent but not in the first. The intersecting fragents are then reoved fro the fragents list. This process continues until no ore intersecting fragents or data set records still exist for this teleedicine relation, and so for the other relations. Start Set of Teleedicine database Relations Set of data records generated by teleedicine queries Execute the fragentation technique for the overlapped and redundant data records Processing Execute the fragentation technique for the nonoverlapped data records End Set of disjoint fragents Figure 2: Data Fragentation Service Architecture Table 2: Overlapped and redundant data records fragentation Algorith Step 1: Set 1 to I; K = F.size() Step 2: Do steps (3-18) until I >F.size() Step 3: Set 1 to J Step 4: Do steps (5-16) until J > F.size() Step 5: If I J and Fi, Fj Є F go to (6) Else, Add 1 to J and go to step (15) Step 6: If Fi Fj Ø do steps (7-14) Else, Add 1 to J and go to step (14) Step 7: Add 1 to K Step 8: Create new fragent Fk= Fi Fjand add it to F Step 9: Create new fragent F k+1 = Fi Fkand add it to F Step 10: Create new fragent F k+2 = Fj- Fk and add it to F Step 11: Delete Fi Step 12: Delete Fj Step 13: Set F + 1 to J Step 14: End IF; Step 15: End IF Step 16: Loop Step 17: Add 1 to I Step 18: Loop Step 19: Add 1 to R Step 20: Loop The Non-overlapped data records fragentation process is shown Table 3. The new derived fragents fro Table 2 and the fragents fro non-overlapping fragents resulted in totally disjoint fragents. The non-overlapped fragents that do not intersect with any other fragent in Table 2 are considered in the final set of the relation disjoint fragents. For exaple, consider that the transactions triggered fro the teleedicine database sites hold queries that require extracting records fro different teleedicine database relations. Let us assue three transactions for patient relation (T1, T2, T3) on sites S1, S2, and S3 respectively. T1 defines Patients who edicated by Doctor #1, T2 defines Patients who took edicine #4, and T3 defines Patients who paid ore than $1000. The data set 1 in patient relation intersects with 13

14 data set 2 in the sae relation. This result in new disjoint fragents: F1, F2 and F3. The Data sets 1 and 2 are oitted. Fragent F1 and data set 3 in the sae relation are intersected. This results in the new disjoint fragents F4, F5, and F6. Then F1 and the data set 3 are oitted, and the next intersection is between F2 and F6. The result is new disjoint fragents; F7, F8, and F9. Then F2 and F6 are oitted. The final disjoint fragents generated fro the transactions over patient relations are: F3, F4, F5, F7, F8, and F9. Table 4 depicts the fragentation process over patient relation. Table 3: Non-overlapped Data Records Fragentation Algorith Step 21: Set 1 to I; K = F.size() Step 22 Do steps (23-34) until I >F.size() Step 23: Set 1 to J Step 24: Do steps (25-32) until J >F.size() Step 25: If I J and Fi, FjЄ Fgo to (26) Else Add 1 to J, go to step (32) Step 26: If Fi Fj= Ø do steps (27-32) Step 27: Add 1 to K Step 28: Create new fragent Fk= Rj- U F Step 29: End IF Step 30: If Fk Ø Add FktoF Step 31: End IF Step 32: Loop Step 33: Add 1 to I Step 34: Loop Table 4: Fragentation process over patient relation Data Set # Site # Transaction # Relation # Generated Record(s)# 1 S1 T1 Patient (1) 39,41,56,63,72,85,97 2 S2 T2 Patient (1) 31,56,72,85,97 Fragent # Site # Transaction # Relation # Generated Record(s)# F1 S1,S2 T1 T2 Patient (1) 56,72,85,97 F2 S1 T1 Patient (1) 39,41,63 F3 S2 T2 Patient (1) 31 Data Set # Site # Transaction # Relation # Generated Record(s)# 3 S3 T3 Patient (1) 17,24,41,63,97 Fragent # Site # Transaction # Relation # Generated Record(s)# F4 S1,S2, (T1 T2) T3 Patient (1) 97 S3 F5 S1,S2 (T1 T2) Patient (1) 56,72,85 F6 S3 T3 Patient (1) 17,24,41,63 Fragent # Site # Transaction # Relation # Generated Record(s)# F7 S1,S3 (T1 T3) Patient (1) 41,63 F8 S1 T1 Patient (1) 39 F9 S3 T3 Patient (1) 17, Clustering Coputing Service The benefit of generating database disjoint fragents can t be copleted unless it enhances the perforance of the WTDS. As the nuber of database sites becoes too large, supporting high syste perforance with consistency and availability constraints becoes crucial. Several techniques are developed for this purpose; one of the consists of clustering web sites. However, grouping sites into clusters is still an open proble with NP-Coplete optial solution [12]. Therefore, developing a near-optial solution for grouping web database sites into clusters helps to eliinate the extra counication costs between the sites during the process of data allocation. Clustering web teleedicine database sites speeds up the process of data allocation by distributing the fragents at the clusters that accoplish benefit allocation rather than distributing the fragents once at all web sites. Thus, the counication costs are iniized and the WTDS perforance is iproved. In this work, we introduce a high speed clustering service based on the least average counication cost between sites. The paraeters used to control the input/output coputations for generating clusters and deterining the set of sites in each are coputed as follows: Counications Cost between sites CC(Si,Sj) = data creation cost + data transission cost between Si,Sj. Counication Cost Range CCR (s/byte) which is deterined by the teleedicine database syste adinistrator. 14

15 Clustering Decision Value (cdv): cdv( S, S ) 1: IF CC( S, S ) CCR i j and 0: IF CC( S, S ) CCR i j (1) i j i j i j Accordingly, if cdv(si,sj) is equal to 1, then the sites Si,Sj are assigned to one cluster, otherwise they are assigned to different clusters. If site Si can be assigned to ore than one cluster, it will be considered for the cluster of the least average counication cost. Based on this clustering service, we develop the clustering algorith shown in Table 5. Table 5: Clustering Algorith Input: Matrix of counication cost between sites CC(Si,Sj) CCR: counication cost range; N: List of WTDS sites; Output: CDV(Sn,Sn) Clustering Decision Values Matrix Step 1: For I = 1, N.size(), do steps (2) - (8) Step 2: For J = 1, N.size(), do steps (3) - (7) Step 3: If I J AND CC(Si,Sj) <= CCR, go to step (4) Else, go to step (5) Step 4: Set 1 to both CDV(Si,Sj) and CDV(Sj,Si), go to step 6 Step 5: Set 0 to both CDV(Si,Sj) and CDV(Sj,Si) Step 6: End IF; Step 7: End For; Step 8: End For; Step 9: Stop The counications cost within and between clusters is required for the fragent allocation phase where fragents are allocated to web clusters and then to their respective web sites. The optial way to copute this cost is to find the cheapest path between the clusters which is an NP-coplete proble [12]. Instead, we use the cluster syetry average counication cost as it has been shown to be fast, reliable and efficient ethod for the coputation of the fragents allocation and replication service in any heterogeneous web teleedicine database environents. To test the validity of our clustering service technique, and based on the assuptions on the counications costs between sites (syetric between any two sites, zero within the sae site, and proportional to the distance between the sites), we collect a real saple of counication costs (ultiple of s/byte) between 12 web sites (hospitals) and present it in Table 6. We then apply the clustering service in our IFCA tool on the counication costs in Table 6 with clustering range of 2.5. Table 6: The counication costs between web sites Site # site1 site2 site3 site4 site5 site6 site7 site8 site9 Site 10 Site 11 Site12 site site site site site site site site site site site site

16 Figure 4 depicts the experiental results that generate the clusters and their respective sites. It is inferred fro the figure that sites 1 and 3 are located in different clusters because the counication costs of the other sites can't atch the counications cost range with the. Moreover, the nuber of clusters is increased as the counication cost between sites increased or the counication cost range becoes sall. Figure 4: Clustering teleedicine web sites 4.3. Data Allocation and Replication Data allocation techniques ai at distributing the database fragents on the web database clusters and their respective sites. We introduce a heuristic fragent allocation and replication coputing service to perfor the processes of fragents allocation in the WTDS. Initially, all fragents are subject for allocation to all clusters that need these fragents at their sites. If the fragent shows positive allocation decision value (i.e. allocation benefit greater than zero) for a specific cluster, then the fragent is allocated to this cluster and tested for allocation at each of its sites, otherwise the fragent is not allocated to this cluster. This fragent is subsequently tested for replication in each cluster of the WTDS. Accordingly, the fragent that shows positive allocation decision value for any WTDS cluster will be allocated at that cluster and then tested for allocation at its sites. Consequently, if the fragent shows positive allocation decision value at any site of cluster that already shows positive allocation decision value, then the fragent is allocated to that site, otherwise, the fragent is not allocated. This process is repeated for all sites in each cluster that shows positive allocation decision value. Figure 5 illustrates the structure of our data allocation and replication technique. In case a fragent shows negative allocation decision value at all clusters, the fragent is allocated to the cluster that holds the least average counications cost, and then to the site that achieve the least counications cost with other 16

17 sites in the current cluster. In order to better understand the coputation of the queries processing cost functions, a atheatical odel will be used to forulate these cost functions. Input Allocation to clusters Allocation to sites Set of teleedicine database disjoint fragents Execute the allocation technique for one cluster Repeat allocation technique for the other clusters Execute the allocation technique for sites of the clusters allocated by fragents Figure 5: Data Allocation and Replication Technique The allocation decision value ADV is coputed as the logical result of the coparison between two copound cost functions coponents; the cost of allocating the fragent to the cluster and the cost of not allocating the fragent to the sae cluster. The cost of allocating the fragent Fi issued by the transaction Tk to the cluster Cj, denoted as CA(Tk,Fi,Cj), is defined as the su of the following sub-costs: retrievals and updates issued by the transaction Tk to the fragent Fi at cluster Cj, storage occupied by the fragent Fi in the cluster Cj, and reote updates and reote counications sent fro reote clusters. The atheatical odel for each sub-cost function is detailed below. Cost of local retrievals issued by the transaction T k to the fragent F i at cluster C j. This cost is deterined by the frequency and cost of retrievals. The frequency of retrievals represents the average nuber ( 0) of retrieval transactions that occur on each fragent at all sites of a certain cluster. The cost of retrievals is the average cost of retrieval transactions that occur on each fragent at all sites of a specific cluster, which is set by the teleedicine database syste adinistrator in tie units (illisecond, icrosecond, etc.) / byte. The cost of local retrievals is coputed as the ultiplication of the average cost of local retrievals for all sites () at cluster C j and the average frequency of local retrievals issued by the transaction T k to the fragent F i for all sites at cluster C j. CLR( T, F, C ) k i j CLR( T,,, ) k Fi C j Sq FREQLR( Tk, Fi, C j, Sq ) (2) q1 q1 Cost of local updates issued by the transaction T k to the fragent F i at cluster C j. This cost is deterined by the frequency and cost of updates. The frequency of updates is coputed as the average nuber of update transactions that occur on each fragent at all sites of a certain cluster. The cost of updates is the average cost of update transactions that occur on each fragent at all sites of a specific cluster, and is deterined by the WTDS adinistrator in tie units per byte. The cost of local updates is coputed as the ultiplication of the average cost of 17

18 local updates for all sites () at cluster C j and the average frequency of local updates issued by the transaction T k to the fragent F i for all sites at cluster C j. CLU ( T, F, C ) k i j CLU ( T,,, ) k Fi C j Sq FREQLU ( Tk, Fi, C j, Sq ) (3) q1 q1 Cost of storage occupied by the fragent Fi in the cluster Cj. This cost is deterined by the storage cost and the fragent size. The storage cost is the cost of storing a data byte at a certain site in a specific cluster. The fragent size is the storage occupied by the fragent in bytes. The cost of storage is coputed as the result of ultiplication of the average storage cost for all sites () at cluster Cj occupied by the fragent Fi and the fragent size. SCP( T, F, C, S ) SCP( T, F, C ) F ( T, F ) (4) k i j k i j q q1 size k i Cost of reote updates issued by the transaction Tk to the fragent Fi sent fro all clusters (n) in the WTDS except the current cluster Cj. This cost is deterined by the cost of local updates and the frequency of reote updates. The frequency of reote updates is the nuber of update transactions that occur on each fragent at reote web clusters in the WTDS. The cost of reote update is the ultiplication of average cost of local updates at cluster C q and the average frequency of reote updates issued by the T k to the fragent F i at each reote cluster in the WTDS. CRU ( T, F, C ) CLU ( T, F, C ) FREQRU ( T, F, C, S ) n n k i q k i j k i q (5) q1, q j q1, q j Update ratio is used in the evaluation of cost of reote counications and represents the estiated percentage of update transactions in the web teleedicine database syste. It is coputed by dividing the nuber of local update frequencies in all WTDS sites by the total nuber of all local retrievals and update frequencies in all sites. Uratio FREQLU ( Tk, Fi, C j, Sq ) FREQLR( Tk, Fi, C j, Sq) FREQLU ( Tk, Fi, C j, Sq) (6) q1 q1 q1 Cost of reote counications issued for the transaction Tk to the fragent Fi sent fro reote clusters (n) in the WTDS. This cost defines the required cost needed for updating the fragent fro reote clusters. It is coputed as the product of the average cost of reote counication, the average frequency of local update, and update ratio. CRC( T, F, C,( S, S )) FREQLU ( T, F, C, S ) n k i q k i j q k i j Uratio q1, q j 1, 1, q1 CRC( T, F, C ) (7) According to the definition of the cost of allocating fragents, and based on the previous allocation sub-costs, the atheatical odel that represents the cost of fragent allocation to a certain cluster in the WTDS is coputed as : 18

19 CA( T, F, C ) CLR( T, F, C ) CLU ( T, F, C ) SCP( T, F, C ) CRU ( T, F, C ) CRC( T, F, C ) (8) k i j k i j k i j k i j k i j k i j To coplete the process of ADV coputation, the cost of not allocating fragent to a certain cluster (the fragent retrieved reotely) should be defined and coputed. The cost of not allocating the fragent Fi issued by the transaction Tk to the cluster Cj, denoted CN(Tk,Fi,Cj), is defined as the su of the following sub-costs: cost of local retrievals issued by the transaction Tk to the fragent Fi at cluster Cj (coputed in equation 2), and the cost of reote retrievals. Retrieval ratio is used in the evaluation of the cost of reote retrievals and represents the estiated percentage of retrieval transactions in the WTDS. The retrieval ratio is coputed by dividing the nuber of local retrieval frequencies in all WTDS sites by the total nuber of all local retrievals and update frequencies in all WTDS sites. Rratio FREQLR( Tk, Fi, C j, Sq ) FREQLR( Tk, Fi, C j, Sq) FREQLU ( Tk, Fi, C j, Sq) q1 q1 q1 (9) Cost of reote retrievals issued by the transaction Tk to the fragent Fi sent fro reote clusters (n) in the WTDS. The cost of reote retrievals is deterined by retrieval ratio, retrieval frequency, and counications cost that represents the counications between all clusters. This cost is coputed as the product of the average cost of counications between all clusters, the average frequency of reote retrieval, and the retrieval ratio. CCC( T, F, C,( S, S )) FREQRR( T, F, C, S )) n n k i q k i q k i j Rraio q1 1, 1, q1, q j 1 CRR( T, F, C ) (10) Based on the previous not allocation sub-costs, the atheatical odel that represents the cost of not allocating fragent to a certain cluster in the WTDS is coputed as: CN ( T, F, C ) CLR( T, F, C ) CRR( T, F, C ) (11) k i j k i j k i j Accordingly, the allocation decision value ADV deterines the allocation status of the fragent at the cluster and its sites. ADV is coputed as a logical value fro the coparison between the cost of allocating the fragent to the cluster CA(Tk,Fi,Cj ) and the cost of not allocating the fragent to the sae cluster CN(Tk,Fi,Cj). ADV ( T, F, C ) 1; CN( T, F, C ) CA( T, F, C ) and 0; CN( T, F, C ) CA( T, F, C ) (12) k i j k i j k i j k i j k i j Therefore, if CN(Tk,Fi,Cj) is greater than or equal to CA(Tk,Fi,Cj), then the allocation status ADV(Tk,Fi,Cj) shows positive allocation benefit and the fragent is allocated at that cluster, otherwise, CN(Tk,Fi,Cj) is less than CA(Tk,Fi,Cj) where the allocation status ADV(Tk,Fi,Cj) shows non-positive allocation benefit, so the fragent is not allocated at that cluster. To foralize the fragent allocation procedures, we developed the allocation and replication algorith shown in Table 7. 19

20 Table 7: Fragents Allocation and Replication Algorith Input: T:List of database transactions; F:List of disjoint fragents (section 3.1); C:List of clusters in the WTDS (section 3.2) Preparation: Step 1: Set 1 to k Step 2: Do steps (3-32) until k >T Step 3: Set 1 to i Step 4: Do steps (5-30) until i >F Step 5: Set False to allocation flag Step 6: Set 1 to j Step 7: Do steps (8-26) until j >C Processing: {Module 1: Coputing costs of allocating and costs of not allocating fragent coponents} {Module 2: Coputing fragent cost of allocation, cost of not allocation, and allocation decision value} {Module 3: Fragent allocation where negative decision value is achieved} Output: The List of fragents that are allocated to each web cluster End. Module 1: Coputing costs of allocating and costs of not allocating fragent coponents, steps 8-17 {Step 8:initialize Cost of Reote Update to 0; initialize Cost of Reote Counication to 0; initialize Cost of Reote Retrieval to 0 Step 9: Set 1 to y Step 10: Do steps (11-17) until y >C Step 11: If y j Then Do steps (12-15); Else Go to step (15) Step 12: Cost of Reote Update(y) = Cost of Reote Update(y -1) + (Cost of Local Update * Average Frequency of Reote Update) Step 13: Cost of Reote Counication(y) = Cost of Reote Counication(y -1) + (Average Cost of Reote Counication * Average Frequency of Local Update * U ratio ) Step 14: Cost of Reote Retrieval(y) = Cost of Reote Retrieval(y -1) + (Cost of Counication between Clusters * Frequency of Reote Retrieval * R ratio ) Step 15: End If Step 16: Add 1 to y Step 17: Loop} Module 2: Copute fragent cost of allocation, cost of not allocation, and allocation decision value. In this odule the total cost of allocation and total cost of not allocation for each fragent will be coputed, as well as the allocation decision value that deterine whether the fragent is allocated to or cancelled fro the cluster, steps {Step 18: Cost of Allocation = Cost of Local Retrieval + Cost of Local Update + Cost of Storage + Cost of Reote Update + Cost of Reote Counication Step 19: Cost of Not allocating = Cost of Local Retrieval + Cost of Reote Retrieval Step 20: Allocation Decision Value= (Cost of Not allocating >= Cost of Allocation) Step 21: If Allocation Decision Value = True Then Go to step (22); otherwise; Go to step (23) Step 22: Allocate the fragent to the current cluster ; Set True to allocation flag; Go to step (24) Step 23: Cancel the fragent fro the current cluster Step 24: End If Step 25: Add 1 to j Step 26: Loop} Module 3: Fragent allocation where negative decision value is achieved. To aintain data availability, fragents that are not allocated to any cluster according to their allocation decision value, will be allocated to the cluster with the least counication cost, steps {Step 27: If allocation flag = False Then Allocate the fragent to the cluster who has the least counication cost Step 28: End If ; Step 29: Add 1 to i Step 30: Loop Step 31: Add 1 to k Step 32: Loop} Once the fragents are allocated and replicated at the web clusters, then the investigation of the benefit of allocating the fragents at all sites of each cluster becoes crucial. To get better WTDS perforance, fragents should be allocated 20

21 over the sites of clusters where a clear allocation benefit is accoplished or where data availability is required. The sae allocation and replication algorith will be applied in this case, taking into consideration the costs between sites in each cluster. For exaple, Consider fragent 1 for allocation in clusters 1, 2, and 3 respectively. Based on Equations 2 through 12 that deterine the allocation decision value, Tables 8 through Table 10 depict the calculations of such process. Cluster # Table 8: Cluster 1 costs of retrieval, update, and storage Cluster # Sites # Cost of retrieval Cost of update Cost of storage Average Table 9: Average # of retrieval, update frequencies, and counication costs Avg. # of Retrievals Average # of Reote Retrievals Avg. # of Updates Avg. # of Reote Updates Avg. cost of co. Avg. cost of reote co Table 10: Coputation of fragent allocation decision value Frag. # Cluster # CA(T k,f i,c j) CN(T k,f i,c j) ADV(T k,f i,c j) Allocation Result CLR CLU SCP CRU CRC CLR CRR 1 (1024 B) CA(T k,f i,c j) < CN(T k,f i,c j) Allocated The allocation decision value for fragent 1 will be coputed in the sae way for the other clusters and for cluster 1 sites (5, 6, and 7) to deterine which cluster/site can allocate the fragent and which one will not Coplexity of Coputation The tie coplexity of our approach is bounded by the tie coplexity of the following incorporated entities: defining fragents, clustering web sites, fragents allocation, coputing average retrieval, and update frequencies. The tie coplexity of each entity is coputed as follows: The coplexity of defining fragents is O(R*(R.size()-1) 2 ), where R is the nuber of relations, and R.size() is the average current nuber of records in each relation. The coplexity of clustering web sites is O((N 2 -N) log 2n), where N is the nuber of sites in the WTDS. The coplexity of fragents allocation is O(R.size()*R*S), where R.size() is the average current nuber of records in each relation, R is the nuber of relations, and S is the nuber of sites. The coplexity of coputing average retrieval and update frequencies iso(f*s*a*a.size()) where F is the nuber of the fragents in the database, S is the nuber of sites, A is the nuber of applications at each site, and A.size() is the average current nuber of records in each application. 21

22 Based on the coplexity coputations of the previous incorporated entities, the tie coplexity of this approach is bounded by: O(R*(R.size()-1) 2 + O((N 2 -N) log 2 n) + R.size()*R*S + F*S*A*A.size()). 5. Experiental Results and Perforance Evaluation To analyze the behavior of its coputing service techniques, we develop an IFCA software tool that is used to validate the teleedicine database syste perforance. This tool is not only experiental but it also supports the use of knowledge extraction and decision aking. In this tool, we test the feasibility of the three proposed coputing services techniques. The experiental results and the perforance evaluation of our approach are discussed in the following sections Evaluation of Fragentation Service The internal validity of our fragentation service technique is tested on ultiple relations of a WTDS. We apply the fragentation coputing service in our IFCA tool on a set of data records obtained fro eight different transactions as (Table 11). Figure 6 depicts the experiental results that generate the disjoint fragents and their respective records. Table 11: Coputation of fragent allocation decision value Transaction Record Transaction Record # # # # 1 1,2,3,7,9 5 18, ,18,20 6 7,8,9,10,13,14,15, ,16 4 1,, ,19 Figure 6: Generating Database fragents Figure 6 shows that our fragentation coputing service satisfies the optial solution of data fragentation in WTDS and generates the iniu nuber of fragents for each relation. This helps in reducing the counications cost and the data records storage, hence iproving WTDS perforance. We now evaluate the fragentation perforance in ters of fragents storage size reduction. Let Fper r represent the fragentation perforance percentage of the relation r, Rdrs r represent the relation data records size generated fro the database queries, and Rfs r represent the relation final fragent data records size, then Fper r is coputed as: Fper r = (Rdrs r - Rfs r ) / Rdrs r (13) For exaple, when the fragentation is perfored for relation 1 on data records sets generated fro the database queries of total size 115 kb and a final nuber of fragents of a total size 33 kb, then the storage size that will be considered for this relation is reduced fro 115 kb to 33 kb with no data redundancy. This action helps in iproving the syste perforance by reducing the data storage size of about 71%. 22

23 Nuber of fragents Nuber of generated clusters To externally validate our approach, we ipleented the two fragentation techniques that are proposed by Ma et al. [8] and Huang et al. [12] respectively and copare their perforance against our approach. Figure 7 shows nuber of fragents against the relation nuber for the three techniques. The results in the figure show that ore large sized fragents are produced by ethods in [8] and [12] due to their fragentation assuptions (See Section 2). The results also show that less nuber of sall size fragents is generated by our coputing service. Therefore, it is clear that our fragentation technique significantly outperfors the approaches proposed by Ma et al. [8] and Huang et al. [12] Fragentation Techniques Coparison (in ters of final nuber of generated fragents) IFCA Ma et. al Huang & Chen Clustering Techniques Coparison IFCA Kuar et. al Franczak et. al Relation Nuber Nuber of Web sites Figure 7: Database fragentation perforances Figure 8: Clustering perforance coparisons 5.2. Evaluation of Clustering Service To evaluate the perforance accoplished by grouping the teleedicine web sites running under our clustering service technique, we introduce a atheatical odel to calculate the perforance gain in ters of the reduced counication costs that can be saved fro clustering web sites. The clustering perforance gain is coputed as the result of the reduced costs of counications divided by the su of counications costs between sites. The reduced counications costs are specifically defined as the difference between the su of costs that are required for each site to counicate with reote sites in the web syste and the su of costs that is required for each cluster to counicate with reote clusters. The perforance evaluation of such atheatical odel is expressed as follows: Let CPE represent (Clustering Perforance Evaluation) and CC(Si,Sj) represent the counication cost between any two sites; Si and Sj in the cluster Ci where i,j = 1,2,3,, n. Let CC(Ci,Cj) represent the counication cost between any two clusters; Ci and Cj where i,j = 1,2,3,, n. CPE can be defined as: n n n n n n CPE CC( Si, S j ) CC( Ci, C j ) CC( Si, S j ) i1 j1 i1 j1 i1 j1 (14) 23

24 For exaple, consider a teleedicine database syste siulated on 10 web sites grouped in 4 clusters, each site counicates with the other 9 sites, and each cluster counicate with the other 3 clusters taking into consideration the counications within the cluster itself. For siplicity, we assue the su of counications costs between sites is s/byte and between clusters is s/byte. By applying CPE (Equation 14), the counications cost is decreased to 74% of the original counications cost. To externally validate our clustering service, we ipleent the clustering techniques by Kuar et al. [13] and Franczak et al. [15] and copare the with our clustering service in ters of the nuber of generated clusters against the nuber of websites (Figure 8). The results in Figure 8 show that our clustering service generates ore clusters for saller nuber of sites; hence it induces less counication costs within the clusters. On the other hand, the other techniques generate less clusters for large nuber of sites, thus, they induce ore counication costs within the clusters. Figure 8 shows that the clustering trend increases with the increase in the nuber of network sites in [13] and ours. In contrast, the nuber of clusters generated by the clustering approach in [15] is less due to their clustering approxiation function that uses natural logarithic function. This in turn results in axiizing the nuber of sites in each cluster which increases the counications cost Evaluation of Fragent Allocation Service To evaluate our fragent allocation coputing, we use a set of disjoint fragents obtained fro our fragentation service, a set of the clusters and their sites generated fro our clustering service, and a set of average nuber of retrievals and updates at each site (Table 12-a). We apply the fragentation service in our IFCA software too using the site costs of storage, retrieval, and update depicted in Table 12-b. the results are shown in Figure 9. Table 12-a: Fragents retrieval and update frequencies Table 12-b: Costs of storage, retrieval, and update 24

25 We propose the following atheatical odel to evaluate the perforance of fragent allocation and replication service. Let IAFS represent the initial size of allocated fragents at site S, n represent the axiu nuber of sites in the current cluster, FAFS represent the final size of allocated fragents at site S, and APEC represents the fragent allocation and replication perforance evaluation at cluster C, then APEC is coputed as: APEC c n s1 IAFS s n s1 FAFS s n s1 IAFS s (15) Figure 9: Fragents allocation at clusters For exaple, assue that 35 fragent allocations with a total size of 91 kb are initially allocated to the WTDS clusters. The final fragent allocation using our approach is 14 with a total size of 27 kb. Therefore, the storage gain of our allocation technique is about 70% which in turn reduces fragents allocation average coputation tie. Our proposed allocation technique is copared with the allocation ethods by Ma et al. [8] and Menon et al. [20]. These two approaches allocate large nubers of fragents to each site and hence consue ore storage which in turn increase the fragents allocation average coputation tie in each cluster. Figure 10 illustrates the effectiveness of our fragent allocation and replication technique in ters of average coputation tie copared to the fragent allocation ethods in [8] and [20]. The figure shows that our fragent allocation and replication technique incurs the least average coputation tie required for fragent allocation. This is because our clustering technique produces ore clusters of sall nuber of web sites which reduces the fragent allocation average coputation tie in each cluster. We Infer fro Figure 8 that the average coputation tie of fragent allocation increases as the nuber of web sites in the cluster increases. Most iportantly, the figure shows that that our technique outperfors the ones in [8] and [20]. 25

26 Average tie (sec) 5.4. Evaluation of Web Servers Load and Network Delay The teleedicine database syste network workload involves queries fro a nuber of database users who access the sae database siultaneously. The aount of data required per second and the tie in which the queries should be processed depend on the database servers running under such network. For siplicity, twelve web sites grouped in 4 clusters are considered for network siulation to evaluate the perforance of a teleedicine database syste. The perforance factors under evaluation are the servers load and the network delay that is siulated in OPNET [23] Fragent Allocation Techniques Coparison IFCA Menon Ma et. al Cluster Nuber Figure 10: Fragents allocation after clustering Figure 11: The WTDS network topology The proposed network siulation for building WTDS is represented by web nodes. Each node is a stack of two 3Co Super Stack II 1100 and two Super stack II 3300 chassis (3C_SSII_1100_3300) with four slots (4s), 52 auto-sensing Ethernet ports (ae52), 48 Ethernet ports (e48), and 3 Gigabit Ethernet ports (ge3). The center node odel is a 3Coswitch, the periphery node odel is internet workstation (S_Int_wkstn), and the link odel is 10BaseT. The center nodes are connected with internet servers (S_Int_server) by 10BaseT. The servers are also connected with a router (Cisco 2514, node 15) by 10BaseT. Figure 11 depicts the network topology for the siulated WTDS which consists of 12 web sites nodes (0, 1, 2, 3, 4, 9, 10, 11, 16, 17, 18, 20) and 4 clusters (6, 13, 14, and 22) connected with internet servers (7 and 8) that are also connected with a router (Cisco 2514, node 15) by 10BaseT. The following subsections evaluate the effect of server load and network delay on the distributed database perforance Server Load The server load deterines the speed of the server in ters of (bits/sec). Figure 12 shows server s load coparison between our approach and the approaches in Kuar et al. [13] and Franczak et al. [15]. The servers are denoted by the cluster legends C1, C2, C3, and C4 respectively. The results of the coparison are drawn fro Figures 11 and 12, and are suarized in Table 13. Table 13 shows that the server load increases as the nuber of represented nodes (sites) 26

27 Load(bits/sec) This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content ay change prior to final publication. Citation inforation: DOI decreases; this is due to load distribution over the nodes assuing that all nodes have the sae capacity. On the other hand, the server load decreases as the nuber of represented nodes increases. The results state that the network is alost balanced throughout processing the web services. However, soe variations are expected due to the possible variations in the nuber of processing transactions on each node. Figure 13 shows the axiu load (bits/sec) on the servers clusters in our proposed clustering technique, and the clustering ethods in [13] and [15]. The results clearly show that our approach outperfors the approaches in [13] and [15] Servers' axiu load Coparison Kuar et. al Franczak et. al IFCA Distance range () Figure 12 : WTDS servers load coparison Figure 13: Servers Max. load for clustering techniques Table 13: WTDS Servers load coparison Server nub. Represent cluster Center node in network topology Represent nodes in network topology Load is well below (bits/sec) 1 C C2 6 0, 1, 2, 3, C3 13 9, 10, C , 17, Network Delay The network delay is the delay caused by the transactions traffic on the web database syste servers. The network delay is defined as the axiu tie (illisecond) required for the network syste to reach the steady state. Figure 14 shows the network delay caused by the WTDS servers represented by the legends (C1, C2, C3, C4). The figure indicates that the web database syste reaches the steady state after illiseconds. It shows that the network delay is less when distributing the web sites over 4 servers copared to the delay consued when all web sites connect 1, 2, or 3 servers. Figure 15 shows the axiu network delay (sec) caused by web servers against distance range for our clustering technique, and the techniques in [13] and [15]. Note that the network delay in our approach is always less than in [13] and [15]; this is due to the better clustering coputations of our technique Threat to Validity Threats to external validity liit the ability to generalize the results of the experients to industrial practice. In order to avoid such threats in evaluation of our approach, we have copared each of our proposed coputing services 27

28 Max. delay (illi sec) techniques; fragentation, clustering and data allocation with two other siilar techniques proposed by different groups of researchers. Each of our proposed techniques is ipleented with two coparable techniques well accepted by the database systes counity. The perforance of our proposed techniques is copared. The results show that our fragentation, clustering and allocation techniques outperfor their counterparts proposed in the literatures Maxiu Network Delay Kuar et. Al Franczak et. Al IFCA Distance range Figure 14: The WTDS network delay Figure 15: Max. Net. delay for clustering techniques 6. Conclusion In this work, we proposed a new approach to proote WTDS perforance. Our approach integrates three enhanced coputing services techniques naely, database fragentation, network sites clustering and fragents allocation. We develop these techniques to solve technical challenges, like distributing data fragents aong ultiple web servers, handling failures, and aking tradeoff between data availability and consistency. We propose an estiation odel to copute counications cost which helps in finding cost-effective data allocation solutions. The novelty of our approach lies in the integration of web database sites clustering as a new coponent of the process of WTDS design in order to iprove perforance and satisfy a certain level of quality in web services. We perfor both external and internal evaluation of our integrated approach. In the internal evaluation, we easure the ipact of using our techniques on WTDS and web service perforance easures like counications cost, response tie and throughput. In the external evaluation, we copare the perforance of our approach to that of other techniques in the literature. The results show that our integrated approach significantly iproves services requireent satisfaction in web systes. This conclusion requires ore investigation and experients. Therefore, as future work we plan to investigate our approach on larger scale networks involving large nuber of sites over the cloud. We will consider applying different types of clustering and introduce search based technique to perfor ore intelligent data redistribution. Finally, we intend to introduce security concerns that need to be addressed over data fragents. 28

29 References [1] Jui-chien Hsieh and Meng-Wei Hsu. A cloud coputing based 12-lead ECG teleedicine service, BMC Medical Inforatics and Decision Making, pp [2] Tahanka, A. and Ra, S. Database Fragentation and Allocation: An Integrated Methodology and Case Study, IEEE Transactions on Systes, Man. and Cybernetics-Part A. Systes and Huans. 1998, V. 28(3), pp [3] Borzeski, L. Optial Partitioning of a Distributed Relational Database for Multistage Decision-Making Support systes, Cybernetics and Systes Research. 1996, V. 2(13), pp [4] Son, J. and Ki, M. An Adaptable Vertical Partitioning Method in Distributed Systes, The Journal of Systes and Software, 2004 V. 73(3), pp [5] Li, S. and Ng, Y. Vertical Fragentation and Allocation in Distributed Deductive Database Systes, The Journal of Inforation Systes, 1997, V. 22(1), pp [6] Agrawal, S.; Narasayya, V. and Yang, B. Integrating Vertical and Horizontal Partitioning into Autoated Physical Database Design, ACM SIGMOD 2004, Paris, France, pp [7] Navathe, S.; Karlapale, K. and Minyoung, R. A ixed fragentation ethodology for initial distributed database design, Journal of Coputer and Software Engineering. 1995, V. 3(4), pp [8] Ma, H.; Scchewe, K. and Wang, Q. Distribution design for higher-order data odels, Data and Knowledge Engineering, 2007, V. 60, pp [9] Yee, W.; Donahoo, M. and Navathe, S. A Fraework for Server Data Fragent Grouping to Iprove Server Scalability, Interittently Synchronized Databases CIKM [10] Jain, A.; Murty, M. and Flynn, P. Data Clustering: A Review, ACM Coputing Surveys. 1999, V. 31(3),pp [11] Lepakshi Goud. Achieving Availability, Elasticity and Reliability of the Data Access in Cloud Coputing, International Journal of Advanced Engineering Sciences and Technologies, Vol. 5(2), [12] Huang, Y. and Chen, J. Fragent Allocation in Distributed Database Design, Journal of Inforation Science and Engineering, 2001 V. 17, pp [13] Kuar, P.; Krishna, P.; Bapi, R. and Kuar, S. Rough Clustering of Sequential Data, Data and Knowledge Engineering, 2007, V.63, pp [14] Voges, K.; Pope, N. and Brown, M. Cluster Analysis of Marketing Data Exaining Online Shopping Orientation: A coparison of K-eans and Rough Clustering Approaches, H.A. Abbass, R.A. Sarker, C.S. Newton (Eds.), Heuristics and Optiization for Knowledge Discovery, Idea Group Publishing, Hershey. 2002, pp [15] Fronczak, A.; Holyst, J.; Jedyank, M. and Sienkiewicz, J. Higher Order Clustering Coefficients, Barabasi-Albert Networks, Physica A: Statistical Mechanics and its Applications. 2002, V. 316(1-4), pp [16] Halkidi, M.; Batistakis, Y. and Vazirgiannis, M. Clustering algoriths and Validity Measures, Proceedings of the SSDBM Conference [17] Ishfaq A.; Karlapale, K. and Kaiso, Y. Evolutionary Algoriths for Allocating Data in Distributed Database Systes, Distributed and Parallel Databases, Kluwer Acadeic Publishers. 2002, v.11, pp [18] Danilowicz, C. and Nguyen, N. Consensus Methods for Solving Inconsistency of Replicated Data in Distributed Systes, Distributed and Parallel Databases. 2003, V.14, pp [19] Costa, R. and Lifschitz, S. Database Allocation Strategies for Parallel BLAST Evaluation on Clusters, Distributed and Parallel Databases, 2003, V.13, pp [20] Menon, S. Allocating Fragents in Distributed Databases, IEEE Transactions on Parallel and Distributed Systes, 2005, Vol. 16-7, pp [21] Daudpota, N. Five Steps to Construct a Model of Data Allocation for Distributed Database Systes, Journal of Intelligent Inforation Systes: Integrating Artificial Intelligence and Database Technologies. 1998, V. 11(2), pp [22] Microsoft SQL Server Available fro: < [Accessed 20 th October, 2012]. [23] MySQL 5.6 available fro: < [Accessed 21th October, 2013]. [24] OPNET IT Guru Acadeic, OPNET Technologies, Inc Available fro:< [Accessed 27th August,, 2013] [25] Hauglid, J.; Ryeng, N. and Norvag, K. DYFRAM: Dynaic Fragentation and Replica Manageent in Distributed Database Systes, Distributed and Parallel Databases, 2010, V. 28, pp [26] Wang, Z.; Li, T.; Xiong, N. and Pan, Y. A Novel Dynaic Network Data Replication Schee Based on Historical Access Record and Proactive Deletion, Journal of Supercoputing Springer DOI: /s z. Published online 19 October

30 [27] Khan, S. and Ahad, I. Replicating Data Objects in Large Distributed Database Systes: An Axioatic Gae Theoretic Mechanis Design Approach, Distributed and Parallel Databases, 2010, V. 28(2-3), pp [28] KhanS, h. and Hoque, L. A New Technique for Database Fragentation in Distributed Systes, International Journal of Coputer Applications V. 5(9), 2010, pp [29] Jagannatha, S.; Mrunalini, M.; Kuar, T. and Kanth, K. Modeling of Mixed Fragentation in Distributed Database Using UML 2.0, IPCSIT, V. 2, 2011, pp [30] Morffi, A.; Gonzalez, C.; Leahieu, W. and Gonzalez, L. SIADBDD: An Integrated Tool to Design Distributed Databases, Revista Facultad de Ingenieria Universidad de Antioquia ISSN (Version ipresa): , No. 47 March, 2009, pp [31] Özsu, M. T. and Valduriez, P. Principles of Distributed Databases, 3rd edition, 2011, Springer, ISBN [32] Mao, G.; Gao, M. and Yao, W. An Algorith for Clustering XML Data Strea Using Sliding Window, The Third International Conference on Advances in Databases, Knowledge, and Data Applications, 2011, pp [33] Paixão, M. P.; Silva, L. and Elias G. Clustering Large-Scale, Distributed Software Coponent Repositories, The Fourth International Conference on Advances in Databases, Knowledge, and Data Applications, 2012, pp [34] Decandia, G.; Hastorun, D.; Japani, M.; Kakulapati, G., Lakshan, A.; Pilchin, A.; Sivasubraanian, S.; Vosshall, P. and Vogels, W. Dynao: Aazon s Highly Available Key-Value Store, ACM Syposiu on Operating Systes Principles, 2007, pp [35] Accessed on 12th Noveber, [36] Chang, F.; Dean, J.; Gheawat, S.; Hsieh, W. C.; Wallach, D. A.; Burrows, M.; Chandra, T.; Fikes, A. and Gruber, R. E. Big Table: a Distributed Storage Syste for Structured Data, Operating Systes Design and Ipleentation, 2006, pp Isail Hababeh holds a PhD degree in Coputer Science fro Leeds Metropolitan University - U.K. He also holds a Master degree in Coputer Science fro Western Michigan University USA, and a Bachelor Degree in Coputer Science fro University of Jordan. Dr. Hababeh research areas of particular interest include, but are not liited to the following: Distributed Databases, Cloud Coputing, Network Security, and Systes Perforance. Issa Khalil received his B.Sc. and M.S. degrees fro Jordan University of Science and Technology in 1994 and 1996, and the PhD degree fro Purdue University, USA, in 2007, all in Coputer Engineering. Iediately thereafter, he joined the College of Inforation Technology (CIT) of the United Arab Eirates University (UAEU) where he was prooted to associate professor in Septeber In June 2013 Khalil joined Qatar Coputing Research Institute (QCRI) as a senior scientist with the cyber security group. Khalil's research interests span the areas of wireless and wire-line counication networks. He is especially interested in security, routing, and perforance of wireless Sensor, Ad Hoc and Mesh networks. Khalil s recent research interests include alware analysis, advanced persistent threats, and ICS/SCADA security. Dr. Khalil served as the technical progra co-chair of the 6th International Conference on Innovations in Inforation Technology, and was appointed as a Technical Progra Coittee eber and reviewer for any international conferences and journals. In June 2011, Khalil was granted the CIT outstanding professor award for outstanding perforance in research, teaching, and service. Abdallah Khreishah received his Ph.D. and M.S. degrees in Electrical and Coputer Engineering fro Purdue University in 2010 and 2006, respectively. Prior to that, he received his B.S. degree with honors fro Jordan University of Science & Technology in In Fall 2012, he joined the ECE departent of New Jersey Institute of Technology as an Assistant Professor. His research spans the areas of network coding, wireless networks, cloud coputing, and network security. 30