Hybrid fuzzy clustering to improve services availability in P2P-based SaaS-cloud

Abstract

Software as a Service is evolving as a leader model for cloud service delivery, enabling service providers to remotely deliver hosted, developed and managed software over the Internet. In parallel, some IT services are moving from traditional Internet services to cloud services based on peer-to-peer technologies. However, the P2P-based cloud is a large-scale, heterogeneous and highly dynamic environment whose performance is highly dependent on its ability to maintain persistent availability of SaaS services. In this paper, we propose an approach for improving SaaS service availability in order to meet service quality requirements and maintain performance in a P2P-Based cloud environment. It is mainly based on a new hybrid clustering mechanism that aims to provide a virtual and optimal infrastructure in order to organize the system peers into distinct clusters represented by virtual nodes forming together a virtual layer. This layer allows not only the distribution of peer providers but also the formation of condensed areas of each service of interest for a set of neighboring peers, which improve the availability probability of services in specific regions. In addition, a service availability measurement model was proposed based on the use of the system’s virtual layer taking into account different entities at different levels. The experimental results show that the proposed approach improves the probability of SaaS service availability and the reliability of the P2P-Cloud system. It responds mainly to the large-scale nature of distributed systems as well as making the best trade-off of maintaining QOS in terms of availability, performance and cost.

Keywords

Cloud computing peer to peer software as a service availability clustering sequential hybridization clustering process thresholded fuzzy c-means hierarchical agglomerative clustering

ï»¿

1. Introduction

Over the last decade, cloud computing has been considered as the pillar of the future Internet generation and one of the most revolutionary concepts of information technology in the 21 ${}^{\text{st}}$ century. It has become a major research topic and an indispensable technology that is increasingly penetrating in all commercial and scientific computing fields. Cloud Computing is a computing paradigm allowing unlimited access to shared hardware and software resources such as: networks, servers, storage, applications and data. In addition, it offers business models for users who can benefit from the provided infrastructure by adopting a service-oriented approach; where services are offered in three models: IaaS (infrastructure as a service), PaaS (platform as a service) and SaaS (software as a service). Consequently, cloud computing has favored the emergence of new services and fundamentally changes the way of how services are built, delivered and used, leading a rapid market evolution [1, 2].

Software as a Service (SaaS) is viewed as the promising IT service model and the leader of cloud service models, which consists of delivering hosted, developed and managed software applications to users via the Internet [3, 4]. These services are provided via pay-per-use or free use, and are accessible from anywhere as long as there is an Internet connection, allowing users to get rid of the burdens of installation and local execution of applications [1, 2]. As a result, the beneficial operational and financial concept of the SaaS model becoming the dominant model for IT service delivery and attracting the interest of researchers and professionals. According to a recent report, the U.S. advanced technology research and advisory company “Gartner” estimates that SaaS services will benefit from significant revenue growth, and that software providers will move from “Cloud-first” to “Cloud only”. This means that “consumption of licensed software will continue to fall, while SaaS and pay-per-use cloud consumption models will continue to grow”. Simultaneously, certain IT services are migrating from traditional Internet services to the cloud based on P2P systems [5, 6] which has emerged as an attractive and promising alternative model offering a significant improvement in the conception of large-scale distributed systems and in the evolution of Internet architectures. Such systems consist of overlay networks formed by a set of heterogeneous, autonomous, dynamic and interconnected peers that voluntarily participate in the network with equivalent functionality, where each node acts as client and server; and collaborate and communicates with others without using an expensive and complex central infrastructure. Depending on the degree of centralization, P2P networks can be divided into: 1) hybrid P2P network, 2) decentralized (pure) P2P networks, 3) structured P2P network with a distributed hash table [7, 8].

The union of these technologies is becoming one of the trendy words of the next generation industry where the use of its services has gained popularity because of its mobility, low cost, storage and computing power, etc. As any technology that requires constant evolution, it must therefore take special attention to its various challenges. Under such P2P-based cloud architecture, the ability of the system to perform its task depends on the reliability of the P2P environment, which depends on the physical and logical architecture of the network and the nodes composing the system. In other words, task execution is allocated to nodes according to client needs. Therefore, the conception of these environments must take into account the physical features of the peers on which the system is operating including: processor speed, memory size, disk capacity … etc.; as well as their logical characteristics that control their dynamic behaviors such as: inactivity, inertia, churn, free-riding … etc.; and network characteristics including: topology, bandwidth, latency, etc. [9]. Therefore, the performance of the system is influenced by these features where any reason for hardware or software failure makes the network’s peer inactive and inaccessible to meet clients’ needs preventing service providers from properly supplying the requested services. P2P-based cloud service providers must achieve identical or even better service levels than traditional cloud-based service in order to ensure the quality of their services, maintain client loyalty and maximize their sustainable financial gains [10, 11].

Recently, the importance of satisfying a permanent and consistent availability of services is becoming one of the most critical factors for the success of Internet-based services and applications, especially services provided via cloud-P2P environments. Traditionally, availability has been limited to local installations of hardware and software resources that businesses and consumers have deployed and maintained. With the advance of cloud services, there is a considerable migration of these resources into the cloud where their availability refers to the availability probability of a system, a network, hardware and software that collectively provide these services during their use. In addition, the service availability in a cloud-P2P system is defined as the property of being accessible and functional at the request of an authorized entity and reflects the system’s ability to meet customer demands, which is also related to the probability of the temporal and functional availability of peer providers [11, 12]. Therefore, ensuring the service availability in such service-oriented systems are more than a simple guarantee of server availability. It is also about providing a reliable and robust communication infrastructure between providers and clients that is able to flexibly and efficiently assure the required functionalities of the services, their availability and other features that may be covered by the term Quality of Cloud Service (QoCS) including: response time, performance, speed, scalability, trust, etc. [13, 14]. Achieving high availability in a P2P-based cloud environment is a complex task - mainly due to the high expectations of end-users who generally do not understand that this requirement is a challenge for system architects. The adoption of mature and powerful technologies to these environments has become necessary in order to effectively analyze and respond to the current challenges and maintain reasonable performance. In this context, data mining has emerged as a relatively new and interdisciplinary field of computing, consisting of the discovery of potentially valid and useful new models from a huge amount of data using several methods at the intersection of artificial intelligence, machine learning, statistics and database systems that have been proposed and evolved over time. It enables the extraction of structured information and useful knowledge for decision making from unstructured or semi-structured web data sources to identify trends and establish relationships between data, predict future behavior, which contributes to reducing personal storage and infrastructure costs as well as providing efficient and secure services to their users [1, 2, 15, 16]. The clustering is considered as one of the most powerful and widely used Data Mining category, the integration of its techniques in such environments has become very important. It has become clear that the future should witness the real power of large amounts of data and its efficient processing on distributed platforms through the use of clustering.

In this paper, we have extended the work proposed by Achache and all in [6] in order to meet the requirements of service availability and maintain performance in a dynamic “hybrid P2P network-based SaaS-Cloud for remote treatment” environment. Since the system proposed by the researchers is built on a peer-to-peer network, its performance is highly dependent on its ability to maintain service availability, which is related to the availability of the network peers. In order to obtain an estimation of service availability from them, we proposed an approach that improves service availability by overcoming the problems of system failures and peer dynamics. The proposed approach consists first of all in analyzing the individual behavior of clients (users/participants) and forming profile data for each of them. This profile data will be used to classify and aggregate the system’s clients according to their profile similarities using a hybrid clustering mechanism that aims to provide a virtual and optimal infrastructure to organize the system’s peers into distinct clusters, where each cluster is represented by a virtual node forming together a virtual layer. As the choice of clustering techniques depends on the nature of the analyzed data, we propose to improve the quality of the clustering of user profiles by combining in ‘a sequential hybridization process’ two well-known “hard” and “soft” clustering algorithms which are “Hierarchical Agglomerative Classification” and “Thresholded Fuzzy C-Means” [17, 18, 19, 20]. This hybridization responds largely to the nature of the manipulated data, allowing each of the client profile vector to be automatically associated with the appropriate number of clusters judged as close without a priori knowing the cluster number and how many clusters it should belong to. Consequently, a virtual layer is formed and added to the architecture of the system proposed in [6]. This layer allows not only the distribution of peer providers but also the formation of condensed areas of each service of interest to a set of neighboring peers, which greatly contributes to improve the availability probability of these services in specific regions in the virtual layer at the request time. In addition, the proposed model for measuring service availability is based primarily on the use of the virtual layer of the system and considers different entities at different levels of the system. To achieve the best performance in a service provision system, the model uses different types of availabilities existing in the literature [21, 22, 23, 24] that are refined to include special considerations in their calculations due to their particular use.

In summary, the main contributions of this work are as follows:

1)
We form the profile data of the system’s clients from a dynamic monitoring and analysis of their behaviors and their interactions with each other and with the services offered by the system. The profile data collected from each client are represented as profile vectors with several attributes where the set of all these vectors generates a matrix of client profile vectors.
2)
We propose a hybrid unsupervised learning approach by combining in a sequential process the Hierarchical Agglomerative Classification algorithm with the Thresholded Fuzzy C-Means algorithm to improve the quality of automatic semi-fuzzy clustering of clients according to their profile vector similarities. The clustering mechanism provides an efficient representation to form the virtual layer of the system resulting in an optimal virtual infrastructure that contributes greatly to improve service availability.
3)
We considered the virtual layer as the kernel for estimating the availability of system services and the peers holding them. With the absence of formal methods for measuring service availability in the P2P networks on which the cloud architecture is built, several availability metrics are redefined and reformulated for use at different levels in the P2P-based SaaS cloud service provisioning system.
4)
We have implemented an experimental framework of our approach and adapted it to the PeerSim simulator. A number of preliminary experiments are being conducted to explore the performance of proposed clustering mechanism on profile data. In addition, the performance of the proposed approach is evaluated using a set of simulations compared to those previously performed in [6] on the basis of a number of performance criteria, particularly average service availability. To the best of our knowledge, we are the first to apply clustering algorithms on client profile data in a P2P-based SaaS cloud service provisioning system for the improvement of service availability.

The remainder of this paper is organized as follows: The literature review on enhancing and maintaining availability in cloud computing is presented in Section 2. Section 3 provides a complete and detailed description of the approach we propose to improve service availability and performance in cloud computing SaaS based on a hybrid P2P network for remote processing. We describe the experimental framework and present the results in Section 4. A comparative study is discussed in Section 5. Finally, Section 6 concludes the paper with future research directions.
2. Literature review

Service availability is considered as one of the key features to maintain QoS in service-oriented distributed environments. Many research efforts and actual studies have been devoted to developing mechanisms to ensure and increase resource availability and performance in distributed environments, especially in cloud computing. The authors in [25] presented a taxonomy to classify research papers and solutions from cloud providers. By adopting this taxonomy proportionally, this section discusses some of the works that aim to improve and maintain availability in cloud computing environments.

Some authors base their availability guarantee approaches on Service Level Agreements (SLAs) between service providers and users that refer to the probability of providing a service according to defined requirements. In [13], the authors proposed a full SLA management solution in cloud computing from a customer-centric perspective in order to maintain the required availability of cloud services. The proposed approach measures and monitors cloud service availability using systems that detect client requests for cloud services and a BlackBox monitoring system that collects data sets (success or failure and latency). In addition, the authors used the adjustment quality analysis (probability distribution) to assess the representation of availability. In [26] the authors proposed an approach for the automatic generation of ranking services for Peer-to-Peer cloud computing systems based on multi-confidence factor ranking which are SLA-monitoring, resource equity algorithm, payment method and tariff scheme. The approach automatically measures the value of QoS parameters, SLA violation, SLA monitoring and SLA penalty. Subsequently, this SLA information is analyzed to generate the ranking list for the user as a cloud provider. This approach can be useful to the user to obtain a trusted service with less cost.

Improving the availability of cloud services in other research papers is ensured by developing new policies or adopting standard methods of replication and scheduling. Saadat and Rahmani [27] proposed a pre-extraction-based dynamic data replication (PDDRA) algorithm. Based on the file access history of the grid sites, PDDRA predicts future needs and pre-extracts the files to the applicant’s grid site, so that the next time that site needs a file, it will be available locally. This will significantly reduce access latency, response time and bandwidth consumption. However, the optimal location of replicas has not been studied in this article. In [28], the authors presented an algorithm that takes into account the number of file requests and the response time to place the replica in the optimal cluster site. As a result, the average execution time of the tasks is reduced to a minimum. The authors in [29] proposed a data replication system in cloud computing built on the use of the frequent pattern extraction algorithm of data mining to identify the popularity of data and generate adaptive threshold for replication, which proved to be a better mechanism to ensure an efficient replication process in the cloud system. System availability and data failure probability will be measured to identify the location where replicated data will be stored. Mohammed et al. in [30] proposed a Light-Weight Data Replication Strategy (LWDRS). The strategy adaptively selects data files to replicate using a light time series technique called Holt linear exponential smoothing (HLES), which analyzes the recent pattern of data file requests and provides predictions for future data requests. Once a replication factor based on file popularity exceeds a specific threshold, the replication signal is launched. The proposed strategy uses heuristic research to dynamically determine the number of replicas as well as the actual data nodes for replication. In [31] the authors proposed a Data replication based scheduling (DRBS) technique that integrates two techniques: data replication and task scheduling to achieve the goal of reducing data access time in the cloud, increasing load balancing and improving the performance of data-intensive applications. In [22], the author clarified the definition of quantitative as well as qualitative availability of end-to-end cloud computing services by considering all components of the service infrastructure. The results show that high availability requires redundancy for components such as network connections and data centers. Other replication strategies and algorithms have been proposed including: Dynamic Data Replication Strategy (D2RS) [32], Modified Fast Spread (MFS) Strategy based on the standard Fast Spread technique [33], Bandwidth Hierarchy Replication (BHR) Strategy [34], Modified BHR Strategy [35], Hierarchical Data Replication (HRM) Model [36], A Cost-Effective Dynamic Data Replication Management Scheme (CDRM) for a cloud storage system with support for Hadoop Distributed File Systems (HDFS) [14], Data Replication Algorithm based on the principle of Maximum Stream Problem at Minimum Cost [37]. These studies show that replication techniques are an essential approach that can be used to improve data accessibility and response time, maximize the rate of successful task execution and minimize bandwidth consumption to ensure the availability and improve the performance of large data applications in the cloud.

Other researchers have considered fault-tolerance techniques to protect cloud computing systems from any kind of hardware or software failure and maintain service availability. In [38], a replication technique called BVACQ (Binary Vote Assignment on Cloud Quorum) has been proposed to preserve data availability and consistency in case of cloud failure. This technique combines replication and fault tolerance mechanisms by considering the fragmentation of the distributed database. In addition, it takes into account the management of transactions, which helps preserve consistency and increase data availability and system reliability. In [40] the authors proposed a fault tolerance mechanism to ensure application availability by addressing software rejuvenation based on virtual machines. A software rejuvenation agent is installed in each virtual machine and monitors applications for software aging and application failures. In this solution, the detection is done at the application level, where replication is done at the VM level, which raises the issue of propagating application failure to waiting VMs. Singh et al. in [41] proposed intelligent failover strategies for cloud data centers using built-in checkpoint algorithms, which include support for load balancing algorithms and multi-level checkpoint. The envisaged failover strategy will run at the application layer and provide high availability for the Platform as a Service (PaaS) functionality of cloud computing. The study [42] investigated the importance of adaptive resource management for fault tolerance in cloud computing applications. The authors extended this concept with an online controller to provide a heuristic algorithm to restore application requirements at runtime in case of failure. Markov chains and queuing network algorithms are used to estimate the availability and performance attributes of a different implementation. The proposed approach increased availability and reduced degradation of system response times compared to traditional static schemes. Other efforts [43, 44] have been devoted to develop middleware and fault-tolerant system architectures able to handle various software failures to ensure the availability and reliability of server applications in a virtualized cloud VDC environment. In addition, some works [45, 46, 47] have developed interesting approaches to achieve adaptive and scalable service availability in Open Stack cloud computing environments and to ensure continuous operation of all cloud services in case of failure. In these works, the researchers attempt to address the weaknesses of existing fault-tolerance approaches by proposing improvements in failover, monitoring and weighting mechanisms, which have helped to minimize average response time and operating costs, improve the degree of required service availability and system performance. Table 1 summarizes the main above works by highlighting their advantages and limitations. Additionally, it presents the main cloud computing performance criteria that have been considered including: (a) availability, (b) scalability, (c) reliability, (d) cost reduction, (e) execution time reduction, (f) bandwidth consumption, (g) load balancing.

By examining these works closely, we realize that the interpretation of availability may differ from one paper to another and from one supplier to another. In addition, most of these techniques had a positive effect on increasing and improving system availability and performance while avoiding frequent outages, however, they are not able to guarantee the desired availability due to their various limitations. In addition to the dynamic volume of data analyzed in distributed environments that requires optimal management, the techniques mentioned above generally allow the production of huge amounts of data circulating in the network, which can limit them to ideally maintain system performance. This large amount of data requires more storage space, which makes it too costly and complex to maintain and manage consistently, unnecessarily increasing management and storage costs. It also requires a global knowledge of the network, which makes them more difficult to implement, limiting the performance of the search for the required data. Besides the processing and execution delay in wide area networks, the difficulty to access data, the increase in response time and network bandwidth limitation, the reduction in priority of fault recovery, the increase in peer inactivity time, they create a very sensitive bottleneck to simultaneous data access in the distributed environment.

In response to these limitations, another research path has advocated a different policy that applies clustering techniques to conveniently and efficiently manage dynamic and heterogeneous cloud resources. Malathy et al. in [48] proposed a resource clustering approach to form clusters based on the identification of the resource usage distribution for a group of nodes with similar resource usage patterns. This

Table 1
Summary of the main papers on improving availability in the cloud through replication, scheduling and fault tolerance mechanisms

Reference	Mechanism	Architecture/algorithm	Advantages	Disadvantages	Main considered performance criteria
					a	b	c	d	e	f	g
[27]	Replication	Pre-fetching based Dynamic Data Replication algorithm (PDDRA)	Improved job execution time, network usage, hit ratio and storage usage. Reduced access latency and bandwidth consumption.	Doesn’t consider the best replica selection.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$
[28]	Replication	Dynamic replica creation and placement algorithm	Increased data availability and effective network usage. Reduced unnecessary replication and placement costs as well as mean job execution time.	Consider only requests number and response time for placing replicas. Don’t include security in the selection of replicas.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$
[29]	Replication and scheduling	Data Replication Model based Frequent Patent Mining	Enhanced data availability, reduced data loss probability, better efficiency and data access time.	High replication cost and low load balancing.	$\surd$	$\surd$	$\surd$	$\times$	$\surd$	$\surd$	$\times$
[30]	Replication	Light-Weight Data Replication (LWDRS)	Optimized Non-Functional QoS and Overall Reliability. High availability and efficient data center access.	High Overhead and low load balancing.	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$	$\times$
[31]	Replication and scheduling	Data replication based scheduling algorithm (DRBS)	Improved response time and load balancing. Low bandwidth. High availability, effectiveness and reliability.	High storage and replication cost.	$\surd$	$\surd$	$\surd$	$\times$	$\surd$	$\surd$	$\times$
[32]	Replication	Dynamic Data Replication Strategy (D2RS)	Enhanced availability and success rate. Minimized network delays and bandwidth usage.	High Replication cost and low load balancing.	$\surd$	$\surd$	$\surd$	$\times$	$\surd$	$\surd$	$\times$
[33]	Replication	Modified Fast Spread (MFS)	Enhanced performance in terms of response time and total bandwidth consumption.	Low reliability and high storage cost.	$\surd$	$\surd$	$\times$	$\times$	$\surd$	$\surd$	$\times$
[34]	Replication	Bandwidth Hierarchy based Replication (BHR)	Reduced data access time and job execution time. Prevents network congestion. Enhanced network efficiency and storage resource usage.	Performs well only when the storage capacity is limited.	$\surd$	$\surd$	$\surd$	$\times$	$\surd$	$\surd$	$\times$
[35]	Replication	Modified Bandwidth Hierarchy based Replication (MBHR)	Reduced mean job execution time, network usage, bandwidth consumption and network traffic. Improved data availability and performance. Minimized data access time and unnecessary replication.	Low load balancing.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$
[36]	Replication	Hierarchical Replication Model (HRM)	High efficiency, accelerated data transmission, improved data availability and accessibility.	Not specified.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$
[37]	Replication	Data Replication quality of service (DRQS) algorithm	Reduced response time and replication cost. Increased Availability, scalability and capabilities.	High storage cost and energy consumption in data centers.	$\surd$	$\surd$	$\surd$	$\times$	$\surd$	$\times$	$\times$
[14]	Replication	Cost-effective dynamic replication management scheme (CDRM)	Cost efficient scheme that can optimize Number of Replicas, Inter-request and Intra-request parallelism. Improved availability, access latency, performance and load balance for real-life applications.	not deal with failed nodes containing deleted file pieces.	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$	$\times$

Table 1, continued
Reference	Mechanism	Architecture/algorithm	Advantages	Disadvantages	Main considered performance criteria
					a	b	c	d	e	f	g
[38]	Replication and Fault Tolerance	Binary Vote Assignment on Cloud Quorum (BVACQ)	Preserve data consistency and system reliability as well as increase the amount of data availability.	High storage cost and low performance due to the replication.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$
[40]	Fault Tolerance	Virtual Machine based Software Rejuvenation (VMSR)	High fault-tolerant prevention, availability and efficiency. Lower management cost for aging application. optimized system performance.	Not adapted to non-deterministic aging profiles and transient failures.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$
[41]	Fault Tolerance	Failover Strategies based on Integrated Checkpointing Algorithms	Provide high availability to the requests of cloud clients in case of failure/recovery of service nodes. Reduction of checkpoint overhead costs.	Don’t incorporate Virtual Machine Monitor rejuvenation. Execution time delays due to random node allocation.	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$	$\surd$
[42]	Fault Tolerance	Adaptive Resource Management for Balancing Availability And Performance in cloud computing	Improved applications availability and performance. Reduced system’s response time even in failure and recovery events.	Don’t support large scale cloud environment.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$

approach aims to improve the performance of the cloud environment in resource utilization. It uses two complementary techniques including: resource utilization histograms to provide statistical information for resource capacities and resource clustering based on Expectation Maximization clustering algorithm “EM” used to obtain compact representation of a set of similar cloudlets to achieve scalability. The experimental results show that the proposed approach is able to provide high accuracy for resource availability and discovery. However, it suffers from low efficiency and high network traffic when the number of clusters becomes large. The paper [49] proposes an architecture based on clustering virtual machines in cloud data centers in terms of various attributes such as RAM, OS type, hardware configuration, etc. Resource clustering applies the K-Means clustering method that helped virtual machines reconfigure themselves and facilitate resource scheduling. It will lead to a better user experience with higher resource availability and scalability. The system proposed in [50] focuses on clustering virtual machines taking into account the size of the requested task and the bandwidth level. The main objective of the proposed virtual machine clustering is to match the task to the appropriate virtual machine for execution with the bandwidth to achieve high availability and maintain high reliability in cloud computing. The proposed algorithm proves its performance offering reduced task execution time with higher efficiency. Wu et al. [51] proposed a classification strategy for cloud service resources. The research presented an improvement to the original naive Bayesian classification algorithm by introducing weights to calculate similarity for different features in the resource dataset. In addition, they implemented a parallel programming model using Hadoop combined with MapReduce to implement the parallelization of the improved classification algorithm. The experimental results show that the proposed strategy is effective in large-scale dynamic cloud environments and offer improved resource classification performance via parallelization. However, this research does not consider applications and only classifies resources of different types. Yoori et al. in [52] proposed a method for dynamic clustering of similar resources by applying the SOM self-organizing maps algorithm and the k-means algorithm in hybrid cloud environments. Based on the resulting clusters, a cost-effective resource recommendation method to cloud users is applied that reflects the characteristics of applications. Experimental results prove the effectiveness of the proposed methods in terms of execution time, availability and resource utilization costs. In [53], a novel technique for optimal resource discovery and dynamic resource allocation is proposed. The proposed technique uses the Modified Hierarchical Agglomerative Clustering (MHAC) algorithm for the discovery of resources and construction of the tree. Subsequently, the resources are allocated using the hybrid artificial bee colony and cuckoo search (HABCCS) algorithm. Here, the artificial bee colony is used to optimize the tree construction path and the cuckoo search is used to modify the artificial bee colony algorithm. From the exciting results, it is clear that this technique allows to improve the availability of desired resources which are allocated in the most efficient way with minimum computation time. Meenakshi et al. [54] proposed an efficient cloud resource provisioning technique using k-means clustering and gray wolf optimization (GWO) partitioning technique. The proposed technique uses GWO for prioritization and k-means clustering is to analyze the QoS metrics to allocate the cloud resources in the best possible way and meet the users’ demands. The experimental results prove the performance of the proposed approach in terms of clustering accuracy effectively providing satisfactory service QoS in terms of availability, load balancing, memory usage and execution time. These research papers have shown that existing resource and service clustering approaches tend to focus on classifying small datasets and their performance decreases in large-scale environments. Table 2 summarizes the above approaches focusing on their benefits and drawbacks. Further, it introduces the main considered performance criteria, namely: (a) availability, (b) scalability, (c) cost reduction, (d) execution time, (e) bandwidth consumption, (f) load balancing, (g) accuracy.

Table 2

Summary of the main papers on improving availability in the cloud through clustering techniques

Reference	Used algorithms	Advantages	Disadvantages	Main considered features
				a	b	c	d	e	f	g
[48]	Expectation Maximization “EM” clustering algorithm	Enhanced availability and fault tolerant. High success and accuracy for resource discovery. Improved scalability and resource utilization.	Low efficiency. High network traffic. Large overhead as the clusters number becomes large.	$\surd$	$\surd$	$\surd$	$\times$	$\times$	$\times$	$\surd$
[49]	K-Means clustering algorithm	High availability and scalability. Improved load balancing, cost and resource utilization. Minimized response time and Bandwidth utilization.	The algorithm requires an appropriate initialization of the clusters number.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$
[50]	Clustering algorithms (not specified)	Maintain high availability, reliability, latency and efficiency. Reduced task execution time and bandwidth consumption.	not specified.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$
[51]	Parallel Bayesian classification algorithm	Improved resource classification accuracy and performance. High availability and efficiency in large-scale dynamic heterogeneous cloud environments.	Doesn’t consider applications and only classifies resources of different types.	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$	$\surd$
[52]	SOM self-organizing maps algorithm and k-means algorithm	Improved availability and efficiency. Reduced execution time. optimized resource utilization costs and VMs energy consumption.	Doesn’t scale to large cloud environments.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$
[53]	Modified Hierarchical Agglomerative Clustering (MHAC) algorithm, hybrid artificial bee colony and cuckoo search (HABCCS) algorithm	Improved availability and efficiency. Minimized cost, memory usage and execution time.	Not specified.	$\surd$	$\surd$	$\surd$	$\surd$	$\surd$	$\times$	$\times$
[54]	K-means clustering and gray wolf optimization (GWO) partitioning technique	High clustering accuracy and efficiency. Enhanced load balancing, availability, memory usage and execution time.	Analysis of cloud workloads is less efficient.	$\surd$	$\times$	$\times$	$\surd$	$\times$	$\surd$	$\surd$

Since the durable availability, accessibility and efficient management of services in distributed environments is of vital importance for both service customers and providers, the integration of clustering techniques due to Data Mining in service-oriented distributed environments such as SaaS models has become necessary in order to maintain reasonable performance [55, 56]. Recently, restricted theoretical and practical studies that favored the integration of clustering techniques in the SaaS model of cloud computing have been developed. These works have been presented in a recent paper [57], where the authors found that the different approaches focus mainly on assessing the quality of SaaS services rather than on improving the SaaS solutions provided. Although the studies cited in this section don’t deal with the aspect of incorporating clustering techniques to improve the availability of SaaS-Cloud services based on P2P networks, we consider that it could be an important step towards a highly available P2P-based cloud services provisioning platform.

3. Proposed approach

In this section, we first give a brief description of the network model for a SaaS-Cloud system based on a H-P2P for remote service processing proposed by Achache et al. in [6] on which our solution will be executed. Then, we describe the metadata structure used to represent the system’s clients. Finally, we present our proposed approach aimed to enhance service availability and maintain a reliable and high-performance P2P network-based SaaS cloud service delivery system.

3.1 System model

Achache et al. in [6] have developed an initiation to distribute the SaaS model for the research field. The system proposed by the authors consists of a new cloud architecture that uses hybrid P2P network (H-P2P) as infrastructure to provide services in SaaS form to peers who don’t have them. It allow peers the use and execute SaaS services that are initially provided by the cloud but hosted and executed by the peers themselves. These services are executed by invocation using Remote Method Invocation (RMI) technology on the peers hosting them. This architecture consists of a distributed dynamic environment that has a minimum infrastructure covering four types of entities including: 1) The administrator responsible for feeding the system with services and their provisioning; 2) Partial Cloud Servers “PCS” are connected on a network layer and are responsible of service indexation. The clients can be: 3) Participant Clients, who host, provide and remotely execute services on behalf of the User Clients, 4) These later, could exploit these services. Together, the PCS and clients, which are geographically distributed, form the set of super peers and normal peers consecutively, hence constituting the H-P2P network. As a result, the normal peers are grouped into clusters managed by super peers forming the real network layer of the system. In addition, the clients exploit the concept of P2P networks to allow a direct connection between them when a service is invoked without any third party intervention.

3.2 Methodology

The proposed approach tries to rally the dynamicity and velocity of peers’ behavior in the real network layer of the service delivery system in order to collect their information. This individual behaviors information forms a profile data as a matrix of client profile vectors. A clustering mechanism uses this matrix and provides an optimal virtual infrastructure reorganizing the peers in the network. In other words, the proposed clustering mechanism performs a new grouping of the network peers besides their clustering in the real network layer. These peers are dispatched, according to their profile’s similarities, in several clusters forming a Virtual Layer (VL). Each cluster is represented by a Virtual Node ( ${VN}_{C_{j}}$ ). This layer enables not only the distribution of the service providers, but also allows the formation of condensed areas of every service that interest a set of peers who host or use them. To achieve this, we propose to improve the quality of the clustering of user profiles by combining, in a sequential hybridization process, two well-known algorithms from the two clustering families “hard” and “fuzzy”; namely “hierarchical ascending classification (HAC)” and “Thresholded Fuzzy C-Means (TFCM)”. Indeed, the proposed hybridization responds widely to the nature of the manipulated data, allowing each of the client profile vector to be automatically associated to the appropriate number of clusters. These later are the most similar without a priori knowledge of the number of clusters and to how many clusters it should belong to. Firstly, the HAC algorithm utilizes the users’ profiles vectors and generates a hard partition that minimizes inter-class inertia and maximizes intra-class inertia. Secondly, this hard partition, with its associated information, is used in the initialization stage of the TFCM algorithm instead of a random initialization. In our approach the TFCM is a basic tool for building a semi-fuzzy hierarchy from user profiles and a complementary algorithm for HAC allowing the assignment of users to clusters judged sufficiently close. We suggest the use of the Z-Score normalization technique to pre-process the variability of the data sets and hence ensuring the accurate functioning of the objective functions and improve the performance of the clustering results. Consequently, the virtual layer produces condensed areas of services hosted by peers belonging to the same cluster, which can greatly improve the probability of availability of these services in specific regions at the time of demand.

In our service delivery system, the proposed service availability measurement is based primarily on the use of the virtual layer to compute the services’ availability probability taking in the account the availability probability of virtual nodes and thus the availability probability of service providers in the real network layer. In order to measure the availability of these entities, we use different types of availability metrics from the literature: “time-based availability, activity-based availability, availability based on the presence of k-of-n”. Our proposal has been structured in three major steps illustrated in the general scheme in Fig. 1 and detailed separately in the next subsections.

Figure 1.

Global scheme of the proposed approach.

3.2.1 Data preparation

3.2.1.1. Client profiles data structure

The information of client profiles in the P2P-based SaaS-cloud service delivery system is maintained in 2D matrix. A client profile is a collection of information that determines its own behavior in the network. We denote $P$ the set of $N$ peers constituting the system, ${P=\{p}_{1},p_{2},\ldots,p_{n}\}$ . The matrix $\textit{VPM}(N\times Q)$ (as shown in the Fig. 2) is a set of $N$ profile vectors ${PV}_{p_{i}}$ and $Q$ variables.

Figure 2.

Client profile vector matrix.

Each profile vector ${PV}_{p_{i}}$ represents the profile of a client $p_{i}$ and consists of a set of $q$ variables, or attributes, describing its behavior. In this work, these attributes are periodically gathered from the system functioning in a fully dynamic environment. Several features characterizing the peer’s behavior in a cloud environment based on a hybrid P2P network for processing remote SaaS services are defined:

Hosted services: this criterion provides an indication of all the services hosted by a peer. Interest towards services: is a $\ll\textit{Client-Service}\gg$ preference matrix inspired from recommender system [58] that provides a preference vector for each client. The preference vector contains assessment values rating a client’s interest in services hosted, requested, or executed accordingly. Geographic location: this attribute describes the peer’s actual spatial. Indeed, geographic location has proven to be an important factor in network optimization in terms of peer-to-peer communication costs. Reputation: Trust is a complex concept that measures the degree of reliability in terms of the uncertainty and subjectivity that one peer can solicit another peer to perform a given task for him. Trust, generally referred to reputation, is seen as a more objective overall measure formed on the basis of information or observations about past behavior [59, 60]. In the proposed system, reputation reflects the degree of peer reliabilities and is calculated by aggregating a set of characteristics.

Time-based features: measure the temporal behavior of peers: (a) connection time, (b) connection duration and (c) temporal availability.

Activity-based features: characterize the level of peer activities and include: (a) individual availability based on the interaction (activity), (b) number of received service requests, (c) number of successful service provisioning requests, (d) number of failed service provisioning requests, (e) number of search requests and finally (f) the average of services hosted during a session.

The above features can vary from one session to another and during one session.

3.2.1.2. Feature normalization

Data normalization is defined as the process of scaling raw data without changing their nature and aim to generate clusters of high quality and improve the performance [61].

Several attributes of the profile data have very large range, for example the value of “number of received service requests” can vary in the range [0, 150]. Therefore, these features are normalized using the Z-score normalization method, which is considered one of the most powerful normalization methods that will give more accurate and efficient clustering results. The Z-score technique is a process of scaling the values of an attribute X based on the mean and standard deviation to ensure that the distribution of features has a mean $\mu=$ 0 and a standard deviation $\sigma=$ 1 (see Eq. (1)).

$\displaystyle Z(\textit{val}_{ij})=\frac{\textit{val}_{ij}-{\mu}_{j}}{{\sigma}% _{j}}$ (1)

3.2.2 Virtual layer formulation

3.2.2.1. Clients’ profiles data clustering

The proposed data clustering mechanism combines the two algorithms HAC and TFCM in order to overcome their limitations and benefit from the advantages of both algorithms. It starts with the application of the HAC algorithm on the data. The HAC algorithm takes as input a user profile matrix formed by a set of profile vectors specified to each system peer $VP=\{VP_{i}(\textit{val}_{1},\textit{val}_{2},\ldots,\textit{val}_{Q}),{VP}_{i% }\varepsilon R^{Q}|i=1,\ldots,N\}$ , where $N$ is the number of network peers represented by profile vectors of dimension $Q$ . Each vector ${VP}_{i}$ is a set of attribute values that describe the interest and behavior of a peer. The output of HAC algorithm is a set of clusters $C=\{C_{m}{\subset}\textit{VPM}|m=1,{\dots},{K}|K\leqslant N\}$ , where $K$ is the number of clusters in the output and each cluster $C_{m}$ is a set of similar profile vectors. In order to know the best partitioning we have added a data storage structure necessary in the automation process of better partitioning detection by the dendrogram. Clusters created by the HAC algorithm are a hard partition of input vectors, which is means: $\forall m,s\in R$ with $1\leqslant\begin{array}[]{c}\mathrm{m,s}\\ m≠s\\ \end{array}\leqslant K$ ; $C_{m}\neq\oslash;C_{s}\neq\oslash$ ; $C_{m}\cap C_{s}=\oslash$ . In addition, the union of all the clusters allows the reproduction of the vector profile matrix, $\coprod^{K}_{m=1}{Cm}=\textit{PVM}$ . These clusters are formed by the following steps (Fig. 3):

Step 1: Considering each of the profile vectors ${VP}_{i}$ , as a single cluster $\{{VP}_{1}\},\{{VP}_{2}\},{\dots},\{{VP}_{N}\}$ . A data storage structure $S$ is defined as a set of triplets ( $r$ , $\textit{Ward's Minimal Distance}_{r}$ , $\textit{Clusters}_{r}$ ), where $r$ is the current cycle, or iteration, reached by the algorithm, $\textit{Ward's Minimal Distance}_{r}$ is the minimum value of inter-class inertia loss values among the values of the distances calculated between all peer profile vectors, $\textit{Cluster}_{r}$ represents the resulting clusters during a cycle ( $r$ ), i.e. $S(0)\leftarrow({0},{0},\{\{{VP}_{1}\},\{{VP}_{2}\},{\dots},\{{VP}_{N}\}\})$ . Measure the similarity between each pair of clusters using Ward’s criterion with Euclidean distance and obtain the distance matrix $D=(d_{ig})N\times N$ . Step 2: Form a new cluster ${VP}_{ig}$ by ${VP}_{i}$ and ${VP}_{g}$ with minimal distance. $S(r)\leftarrow(r++,$ Ward’s Minimal Distance ${}_{r}$ , $\textit{Clusters}_{r}$ ). Step 3: Update the distance matrix by replacing the two clusters ${VP}_{i}$ and ${VP}_{g}$ by the new cluster ${VP}_{ig}$ and calculating its distance from the other clusters. Step 4: Repeat steps 2 and 3 until all objects belong to the same cluster. Step 5: Retrieve the Data Storage structure and use it in the process of automatic detection of better partitioning of the dataset which is carried out as follows:

•
Calculate the hops between each successive pairs of partitions $P_{i}$ and $P_{i+1}$ which is equal to the difference between the minimum inter-class inertia losses values obtained in successive levels of aggregation.

$\displaystyle\textit{Hop}_{P_{i},P_{i+1}}=|\textit{Ward's Minimal Distance}_{P% _{i+1}}-\textit{Ward's Minimal Distance}_{P_{i}}|$ (2)
•
Search for the largest hop among the calculated hops between each two successive partitions: ${\textit{Max}}_{\textit{hop}_{P_{i},P_{i+1}}}$
•
Since we are trying to minimize the loss of inter-class inertia, the final partition to be considered is the one that precedes the largest hop-loss. Retrieve the partition $P_{i}$ and consider it as the final dispersion.

Figure 3.
Hierarchical ascending clustering algorithm with Ward’s criterion.

Based on this partition, information and data required for initialization of the TFCM algorithm can be extracted. Considering the user profile matrix VPM presented previously $P_{\textit{HAC}}$ is the resulting partition of the HAC algorithm with ${P_{\textit{HAC}}}=\{C_{1},C_{2},\ldots,C_{k}\}$ . The Thresholded Fuzzy C-Means algorithm proceeds as follows (Fig. 4):

Step 1(Initialization step). First, the number of clusters $K$ is set by attributing a value equal to the number of clusters of the resulting partition of the HAC algorithm, $K=|P_{j\textit{HAC}}|$ . Secondly, we initialize the membership matrix $U^{r=0}$ based on clusters generated by HAC where each user’s profile vector $VP_{i}$ will be assigned a degree of membership of 1 to the cluster to which it belongs and a 0 for other clusters:

$\displaystyle u_{m}(P_{i})=u_{im}=\left\{\begin{array}[]{ll}1;&\text{if }P_{i}% \in\ C_{m}\\ 0;&\text{otherwise}\\ \end{array}\right\};\text{where}:1\leqslant i\leqslant N,1\leqslant m\leqslant K.$ (3)

The result will be a $K\times N$ binary matrix:

$\displaystyle U^{r=0}=\left(\begin{array}[]{ccc}u_{ij}&\cdots&u_{iK}\\ \vdots&\ddots&\vdots\\ u_{Nj}&\cdots&u_{NK}\\ \end{array}\right)$

Finally, we arbitrarily assign a real value to the fuzziness degree parameter “ $m$ ” knowing that it is greater than 1 ( $m>1$ ) and a value between 0 and 1 to the termination criterion $\varepsilon(\varepsilon\in[0,1])$ . Step 2. At each iteration $r$ the centroid $v_{j}$ of each cluster $C_{j}$ is calculated with $U^{r}$ by applying the following equation:

$\displaystyle v_{j}=\frac{\sum^{N}_{i=1}{{{(u}_{ij})}^{m}{VP}_{i}}}{\sum^{N}_% {i=1}{{{(u}_{ij})}^{m}}}$ (4)

Step 3. Updating the membership matrix $U^{r}$ , $U^{r+1}$ using the following equation:

$\displaystyle u_{ij}={\left[\sum^{K}_{l=1}{{\left(\frac{||{VP}_{i}-v_{j}||}{||% {VP}_{i}-v_{l}||}\right)}^{\frac{2}{m-1}}}\right]}^{-1}$ (5)

Step 4. If $||U^{r+1}-U^{r}||\leqslant\varepsilon$ , go to step 5, otherwise return to step 2. Step 5. Formation of Soft clusters. The algorithm in this step must be able to form the final semi-fuzzy partition where the members of each cluster should have a sufficient degree of membership. This is done, proceed the following steps:

•
Threshold step: Evaluate the degree of membership $u_{ij}$ of each user profile vector ${PV}_{i}$ to a cluster $j$ in relation to the threshold $\tau$ defined as the ratio of the number of clusters:

$\displaystyle\tau=\frac{1}{K},\text{where }k\text{ is the number of clusters.}$ (6) $\displaystyle\left\{\begin{array}[]{l}\text{if }u_{ij}\geqslant\tau;\text{add % user profile vector }i\text{ to cluster }j.\text{ such as }1\leqslant i% \leqslant n\text{ and }1\leqslant j\leqslant c.\\ ;\text{otherwise}.\\ \end{array}\right.$ (7)
•
Normalization step: this step allows to calculate the new degrees of membership ${u_{ik}}^{}$ using the following equation:

$\displaystyle{u_{ij}}^{}=\frac{u_{ij}}{\sum^{k}_{j=1}{u_{ij}}};\text{where}:1% \leqslant i\leqslant n\text{ and }1\leqslant j\leqslant k$ (8)

Figure 4.
Modified thresholded fuzzy C-means.

As a continuation of the HAC* algorithm, the TFCM algorithm allows first of all to add a membership degree factor $u_{ij}$ to each user profile vector ${VP}_{i}$ . It also makes it possible to associate each profile vector with clusters to which it has a sufficient degree of membership. By using a membership function $u_{ij}$ , TFCM measures the degree of membership of the user profile vector ${VP}_{i}$ to each cluster $C_{m}{\subseteq}C$ and its value varies between zero and one $u_{ij}\varepsilon[0,1]$ . In addition, the sum of the degrees of membership for each user profile vector ${VP}_{i}$ must be equal to one $\sum^{k}_{j=1}{u_{ij}}=1$ . The result generated by TFCM is a fuzzy partition FPA of the set of profile vectors $C=\{C_{m}\subset\textit{VPM}|m\epsilon K\leqslant N\}$ , where each vector can belong to more than one cluster among the $k$ resulting clusters but not necessarily to all of them. The union of all the clusters allows the reproduction of the vector profile matrix, which is means: $\forall m$ , $s\in R$ with $1\leqslant\begin{array}[]{c}\mathrm{m,s}\\ m\neq s\\ \end{array}\leqslant K$ ; $C_{m}\neq\oslash$ ; $C_{s}\neq\oslash$ ; $\coprod^{K}_{m=1}{Cm}=\textit{PVM}$ while their intersections do not necessarily give the empty set. The pseudocode of the proposed clustering mechanism is presented in Algorithm 1. This code illustrates the proposed hybrid clustering algorithm including HAC and TFCM.

3.2.2.2. Virtual nodes and virtual layer formation

The set of clusters obtained in the final partition is denoted by $C=\{C_{1},C_{2},\dots,C_{K}\}$ , where each cluster $C_{m}$ is made up of $X$ vectors of similar profiles. After the formation of these clusters, a virtual node ${VN}_{C_{m}}$ will be formed in each cluster ${C}_{m}$ and considered as its representative. In other word, the virtual node is an abstract entity consisting of a vector representing a specific cluster. Each virtual node is generated from an aggregation process that carried out by averaging the values of the $X$ , $(X\leqslant P$ ) profile vectors forming the cluster. The result of this aggregation will be a set of $K$ vectors where each of them represents a virtual node, and formed according to the following equation:

$\displaystyle{VN}_{C}=\left\{{{VN}_{C_{m}}(VN}_{C_{m1}},{VN}_{C_{m2}},\ldots{% VN}_{C_{mq}});1\leqslant m\leqslant K;\phantom{\frac{\sum^{X}_{l=1}{{{(PV}_{l}% )}_{j}}}{X_{C_{m}}}}\right.$ (9) $\displaystyle\quad\left.{VN}_{C_{mj}}=\frac{\sum^{X}_{l=1}{{{(PV}_{l})}_{j}}}{% X_{C_{m}}}\ ;1\leqslant j\leqslant Q\right\}$

Therefore, the fuzzy client partition and the resulting virtual nodes together form a virtual layer that will be added to the architecture proposed by Achache et al. [6].

As mentioned above, the virtual layer allows the formation of condensed locations of each service, which increases the probability of its availability at the time of its request in a specific region. In addition, since the geographic location of peers is one of the characteristics taken into account during the clustering process, peers from different and similar geographic areas and time zones tend to be grouped together in the same cluster. This will have a positive impact on service availability and system performance. On the one hand, geographical proximity between neighboring peers increases the probability of finding close nodes that respond to service demands. On the other hand, the geographical diversity of neighboring peers is advantageous in terms of availability allowing services to be hosted and executed by peers in different geographical locations and to have a presence in several physical areas.

Once a system client performs a search for a service, the search is initially carried out in the virtual layer and rather than using all system users. Therefore, the requestor client contacts only the providers that are part of their virtual cluster. However, if this search fails it will be launched in the “Partial Cloud Servers” layer. The requestor client sends a service search request to his own Partial Cloud Server in order to locate the service and search for a provider. As a result, this will increase the probability of availability of the required services, reduce the search space and improve the response time.
3.2.3 Availability measurement

The proposed availability measurement model refers to the efficient calculation and evaluation of availability in a dynamic network. The basic idea of our service availability assessment is mainly based on the use of the virtual layer of the system. In other terms, the service availability evaluation relies on the calculation of the availability of some entities including the availability of the network peers as well as the availability of the virtual nodes. To achieve the best performances in a service delivery system, different types of availability would be important at different levels of the system. These availabilities maintain the basic core of their traditional definitions, therefore they may appear similar to those typically used in P2P networks but are refined to include special considerations in their calculations due to their particular use:

Service Provider Availability: as the peer providers constitute the unit that performs the task of hosting and providing services to service requesters, increasing their availability probability is important for the service provision system. The Temporal Availability Probability of a service provider ${\textit{TAP}}_{p(i)}$ in a service system, using time-based availability formulas, is simply defined as:

$\displaystyle\textit{TAP}_{p(i)}=\frac{\textit{Up Time}}{\textit{Total % Observation Time}}=\frac{\sum^{S}_{z=1}{t_{p}(i)}}{T}$ (10)

where $t_{p}(i)$ the duration of the session of peer $P$ , $S$ is the number of sessions in which $P$ is connected, $T$ is the overall system uptime. The Activity-based Availability Probability of a service provider $\textit{AAP}_{p(i)}$ for a time interval in a service system reflects the rate of its ability to provide the service correctly. It is determined based on a ratio of its success in providing the requested services to the total number of service requests received by all peer requesters and is calculated as follows:

$\displaystyle\textit{AAP}_{p(i)}=\frac{\textit{Total Success}}{\textit{Total % Request Number}}=\frac{\sum^{R}_{r=1}{S(r)}}{R}=1-\frac{\sum^{R}_{r=1}{F(r)}}{R}$ (11)

where $R$ is the number of requests that the peer provider $p(i)$ has received during a specific time interval, $S(r)$ is the success of the provider in correctly providing the requested service ( $r$ ), $F(r)$ is the failure of the provider to correctly provide the requested service ( $r$ ). Virtual Node Availability: As mentioned previously, the main objective of the virtual layer is to maintain the service availability in the cloud-P2P service delivery system. This is based primarily on virtual nodes. The availability probability of virtual nodes in a given time interval is obtained based on the k-of-n availability metric. A virtual node ${VN}_{m}$ is said available if a subset of $k$ peers from the set of $n$ peers forming the cluster $m$ has a sufficient availability probability. Usually, this measure is based on a line function that takes two possible values, either 1 when all $k$ peers are connected at the measuring time interval, otherwise 0. However, our $k-of-n$ availability measure allows us to calculate the degree of availability probability of an object and a set by adopting the same concept. In other words, the availability probability of a virtual node depends on the degree of availability probability of the $k$ peers among the $n$ peers that constitute its virtual cluster. The method begins by sorting the set of peers constituting the cluster $m$ in descending order according to their availability probabilities, where the peer with the highest (/lowest) availability probability value will be located at the top of the formed list. Then, the first $k$ peers (/the last $k$ peers) with the largest (/smallest) availability values will be selected and considered to calculate the maximum (/minimum) availability probability of the virtual node representing the cluster $m$ . The availability probability of a virtual node is given by the arithmetic mean of the availability probabilities of the selected $k$ peers, i.e. their Time-Based or Activity-Based Availability Probabilities and it defined as follow:

$\displaystyle{AP}_{{VN}_{m}}=\frac{\sum^{k}_{i=1}{{AP}_{p_{i}}}}{k}$ (12) $\displaystyle\textit{TAP}^{k-of-n}_{{VN}_{m}}=\frac{\sum^{k}_{i=1}{\textit{TAP% }_{p(i)}}}{k}$ (13) $\displaystyle\textit{AAP}^{k-of-n}_{{VN}_{m}}=\frac{\sum^{k}_{i=1}{\textit{AAP% }_{p(i)}}}{k}$ (14)

where $\textit{TAP}^{k-of-n}_{{VN}_{m}}$ and $\textit{TAP}^{k-of-n}_{{VN}_{m}}$ reflects respectively the time-based and activity-based availability probabilities of a virtual node m and its values are between the minimum and maximum values of the time-based and activity-based availability probabilities of the $k$ peers; $\textit{TAP}^{k-of-n}_{{VN}_{m}}\epsilon\linebreak\left[{\textit{Min}_{\textit% {TAP}_{p(i)}}},\textit{Max}_{{\textit{TAP}_{p(i)}}}\right]$ , $\textit{AAP}^{k-of-n}_{{VN}_{m}}\epsilon\left[{\textit{Min}_{\textit{AAP}_{p(i% )}}},\textit{Max}_{{\textit{AAP}_{p(i)}}}\right]$ . Based on the obtained value of this metric, we can indicate if the virtual node is available or not. This is done by defining a line function $O_{VN(m)}(t)$ which takes a value of 1 if the availability probability of the virtual node is greater than or equal to a prefixed threshold, otherwise it takes a value of 0 and is formulated as follows:

$\displaystyle O_{{VN}_{m}}(T)=\left\{\begin{array}[]{l}{\text{If }\textit{TAP}% ^{k-of-n}_{{VN}_{m}}\geqslant\beta\text{ or }\textit{TAP}^{k-of-n}_{{VN}_{m}}}% \geqslant\beta;\text{Then 1}.\\ \text{Otherwise};0.\\ \end{array}\right.$ (15)

Availability of the service provision system: The availability of a system during a measuring period reflects its ability to guarantee its main objective, which is the maintenance of an adequate rate of provision of the requested services under all constraints. This clearly shows that service availability and network peer availability is important to us when we want to know the overall state of the system at a given time. We identify it in a service provisioning network as the service provisioning system availability. In a time-based approach, availability defines the probability that the service delivery system is in an active state during a given time interval that is influenced by the temporal availability of the peers constituting the system. The Time-based Availability Probability of the Service Provision System can be expressed as follows:

$\displaystyle\textit{TAP}_{\textit{SPS}}=\frac{1}{N}\sum^{N}_{i=1}{\textit{TAP% }_{p(i)}}$ (16)

In the case of the activity-based approach, availability determines the probability that the service provision system will be functional for a specified period of time. This is reflected in the average ability of peers to be functional in terms of the effective delivery of the requested services over the same period of time. The calculation of Activity-based Availability Probability of the Service Provision System is similar to the previous one.

$\displaystyle\textit{AAP}_{\textit{SPS}}=\frac{1}{N}\sum^{N}_{i=1}{\textit{AAP% }_{p(i)}}$ (17)

In the case of presence-based availability, the availability probability of the service provisioning system would be equal to the average availability probability of the virtual nodes representing the different clusters, which is also equal to the average availability probabilities of the different clusters in the network with $n$ peers, where $k$ of them is sufficient for full functionality:

$\displaystyle{\textit{PAP}^{k-of-n}_{\textit{SPS}}}=\frac{1}{N}\sum^{N}_{i=1}{% {\textit{PAP}^{k-of-n}_{{VN}_{(i)}}}}$ (18)

Service availability: In the cloud-P2P service delivery system, service availability means that the service must be usable by clients when they need it and during the necessary time to provide the service effectively, and that it must be satisfactory for peer requesters. This requires the availability of at least one peer provider that hosts the required service at the time of the request. The assessment of Service Availability Probability during a given time interval is based on the virtual layer of the system where a set of peers is dispersed over a number of clusters represented by virtual nodes. It is expressed as the ratio between the sum of the degree of probability availability of each virtual node weighted by the value of availability of each of the peers possessing the service among the peers composing it and the sum of the degree of probability availability of each virtual node weighted by the maximum value of probability availability that each of its peers possessing the service can reach which logically takes the value of 1. The Service Availability Probability is obtained as follows:

$\displaystyle\textit{SAP}_{S}=\frac{\sum^{K}_{m=1}{{\sum^{n}_{i=1}{{AP}_{{% \textit{Provider}_{S}}_{im}}}}*{AP}_{{VN}_{m}}}}{\sum^{K}_{m=1}{{{\textit{Max}% \_AP}_{{\textit{Provider}_{S}}_{im}}}*{AP}_{{VN}_{m}}}}$ (19) $\displaystyle\textit{TSAP}_{S}=\frac{\sum^{K}_{m=1}{{\sum^{n}_{i=1}{\textit{% TAP}_{{\textit{Provider}_{S}}_{im}}}}*{\textit{TAP}}_{{VN}_{m}}}}{\sum^{K}_{m=% 1}{{{\textit{Max\_TAP}}_{{\textit{Provider}_{S}}_{im}}}*T{AP}_{{VN}_{m}}}}$ (20) $\displaystyle\textit{ASPA}_{{}_{S}}=\frac{\sum^{K}_{m=1}{{\sum^{n}_{i=1}{% \textit{AAP}_{{\textit{Provider}_{S}}_{im}}}}*\textit{AAP}_{{VN}_{m}}}}{\sum^{% K}_{m=1}{{\textit{Max\_AAP}_{{\textit{Provider}_{S}}_{im}}}*\textit{AAP}_{{VN}% _{m}}}}$ (21)

where $\textit{TAP}_{{\textit{Provider}_{S}}_{im}}$ and $\textit{AAP}_{{\textit{Provider}_{S}}_{im}}$ are the time-based and activity-based availability probability of the peer provider $i$ of service $S$ belonging to cluster $m$ . $\textit{TAP}_{{VN}_{m}}$ and $\textit{AAP}_{{VN}_{m}}$ are the time-based and activity-based availability probability of the virtual node $m$ . $\textit{Max\_TAP}_{{\textit{Provider}_{S}}_{im}}$ and $\textit{Max\_AAP}_{{\textit{Provider}_{S}}_{im}}$ are the values of the maximum time-based and activity-based availability probability that a peer can reach and logically it is equal to 1. According to this formula, the availability probability of a service is increased in parallel with the fact that the service is hosted by many peers and that these peers have a high degree of availability probability and belong to several clusters whose virtual nodes have high availability probability.

The service availability in our approach is conditioned by the availability of other entities in a closed cycle as shown in Fig. 5. In other words, a service is available once the service provision system on which it runs is available. Furthermore, a system is available if there is at least one available peer belonging to one or more clusters represented by Virtual Nodes in addition to communication links. Moreover, a Virtual Node is available if there are at least $k$ peers available among its $n$ peers. A peer is considered available if it is connected to the system and is able to efficiently perform an allocated task i.e. correctly provide the requested services. Consequently, a service is available when a provider node is available and accessible at the time of the service request that it needs to process it efficiently.

Figure 5.

Availability cycle.

4. Evaluation and results

This section starts with a brief description of the experimental environment and methodology that we used. Then, we analyze the obtained simulation results in order to measure the impact of the proposed approach on service availability and system performance.

4.1 Experimental setting

We have implemented a prototype of a hybrid P2P network-based SaaS-Cloud for remote treatment that relies on a hybrid fuzzy clustering mechanism for client profiles in order to improve the availability of the services that hold them. The prototype remains the core of the implementation carried out in [6] while taking into account other considerations. In order to study the performance and the impact of the proposed approach on service availability, our experiments are implemented according to three different scenarios, namely:

Scenario 1: in which we implemented a Hybrid P2P-based SaaS cloud service provisioning system without using a clustering mechanism, allowing the evaluation of the proposed approach’s performance and the formation of cache data to be used during network initialization in a different scenario. Scenario 2: in which we implemented a P2P-based SaaS cloud service provisioning system that uses different clustering algorithms (FCM, HAC) applied separately on the overlay network. Scenario 3: where, we have implemented a P2P-based SaaS cloud service provisioning system that uses the proposed hybrid clustering algorithm (HAC-TFCM) as a clustering mechanism.

The proposed approach is compared with the system proposed by Achache et al. in [6] in order to study the service availability factor in both architectures. In addition, the proposed clustering approach was evaluated according to different algorithms to measure its quality as well as its impact on service availability and system performance. The PeerSim simulator, [62, 63], is used as an experimental framework and the results are studied using the metrics described in the next section.

Table 3
Configuration settings

Network size	10000, 20000, 30000, 40000, 50000 nodes
Number of partial cloud servers	30
Maximum number of neighbors for each partial server cloud	5
TTL (time to live)	7
Communication technique between partial cloud servers	Flooding
Number of peers to be added to/removed from network	Varies between 10% and 20% of the network size. For
	deletion, peers are chosen randomly or according to their
	reputation.

A large number of dynamic simulations using the cycle-based PeerSim model are performed to evaluate our prototype. The simulation begins by reading the configuration file consisting of a simple ASCII text file, essentially composed of key-value pairs and contains all simulation parameters concerning all objects involved in the experiment including the definition of the network variables, the declaration of the node structures, protocols, Dynamics and observers used [62, 63]. For more clarity, our experiments used variable test configurations by combining the parameters presented in Table 3. Note that all simulation results are performed on a PC DELL G5 15 5587 with a 2.2 GHz; Turbo 4.1 GHz/9 Intel Core i7 processor, 32 GB memory and a 1To HDD $+$ 256 SSD disk. In addition, we have implemented a set of classes that describe the behavior of the main actors in the different layers of the system:

The administrator: represented by a Control class that simulates its behavior in the cloud layer, accessible by all peers and uses a “Global Service List” as a data structure. The latter contains all the information related to the system’s services (including: their names, categories, inputs and outputs) and is used by peers as a reference for selecting the services to host or request. Besides the tasks allocated to the administrator in the system proposed in [6] including the feeding of the system with services as well as their management, this node will be responsible for the clustering process scheduling in the system implementing the proposed. This is done by specifying the cycles or moments in which the task of node clustering is carried out. Partial Cloud Servers: represented by a class that has the indexing, search modules and tools for communicating with their neighbors at the “Partial Cloud Servers” layer. This class implements the Control interface with a main method “execute ()” to inform each client of his PCS address. It uses “Partial list of hosted services” about the services hosted locally by its peers. Clients: represented by a class containing the necessary modules to perform various tasks including: communicating with their PCS and virtual neighbors, hosting, searching, requesting services and sending results. This class implements the CDProtocol interface that inherits from the Protocol class and has the nextCycle () method where the various tasks that a peer must perform during its execution are implemented including: 1) hosting services, 2) searching for services, 3) deleting services and 4) the peer latency. These scenarios are provided by the system with a probability equal to 0.25 for each one. This class uses the following data structures:

•

Local list of hosted services which contains all hosted services specific to each peer.

•

Partial list of virtual neighbors which contains all the neighboring nodes in the virtual layer. This list will be used in the first place as a basis for searching a provider for a required service and in case the search fails the peer switches to searching via its Partial Cloud server.

Based on the configuration file, the cycle-driven simulation engine is loaded to configure the network by initializing the nodes and network protocols as instances of the classes implementing one or more interfaces. They are created by the “clone ()” method of the “Node” class in a single instance using the object constructor. The initialization phase is carried out by control objects whose execution is programmed only at the beginning of each experiment. After initialization, components (protocols and commands) are called by the cycle-driven engine once per cycle. However, other components are configured to run only in certain cycles until the end of the simulation [62, 63]. At the end, the collected statistical data by the observer components during the simulation is formatted and sent to a standard output for specific tasks, i.e. analysis or storage. Our observer components inherit from the “Observer” class of Peersim and implement a modified Analyze() method.

4.2 Evaluation criteria

In order to evaluate the performance of our system as well as the impact of the proposed approach on the availability of services, we used different evaluation criteria commonly used to estimate the validity of the hybrid fuzzy clustering mechanism and other criteria to measure the degree of services availability quality. These criteria are:

4.2.1 Clustering evaluation criteria

As the actual structure of the input experimental data is unknown where data class labels and external information are not available, the validation of the fuzzy clustering approach requires the use of internal measures that rely only on the information contained in the analyzed data set [64, 65]. We present below some of the most common internal measures used in this paper to estimate the results quality of the hybrid fuzzy clustering mechanism compared to traditional hard clustering techniques and to determine its impact on service availability and system performance.

Partition Coefficient (PC): is considered as the first validity index proposed by Bezdek in [66] to measure the amount of “overlapping” between clusters. The $P C$ index is mainly based on minimizing the pairwise fuzzy intersection in the U partition matrix. It is defined as follows:

$\displaystyle PC(_{c})=\frac{1}{N}\sum^{c}_{i=1}{\sum^{N}_{j=1}{{(u_{ij})}^{2}}}$ (22)

where $N$ is the number of elements, $u_{ij}$ is the memberships of data point $j$ in cluster $i$ . The $P C$ index values vary in the range of [1/c, 1] and as the index is closer to unity, the quality of clustering is better. Partition Entropy (PE): is an index based on the simple entropy, [66], measuring the fuzziness of the cluster partition only. It is defined as follows:

$\displaystyle PE=-\frac{1}{N}\left(\sum^{c}_{i=1}{\sum^{N}_{j=1}{{u_{ij}{\log}% _{b}(u_{ij})}}}\right)$ (23)

$P E$ is calculated for values of $N$ greater than 1 and its values are obtained in the range [0, $\log_{b}(c)$ ]. In contrast to the $P C$ index, smaller $P E$ values show the presence of well-separated clusters while cluster structures become fuzzy if the $P E$ values approach the upper limit of the range. Partition Index (SC): is the ratio of the sum of cluster compactness and cluster separation. It is a sum of individual cluster validity measures, normalized by dividing on the fuzzy cardinality of each cluster [67]:

$\displaystyle SC(c)=\sum^{c}_{i=1}{\frac{\sum^{N}_{j=1}{{(u_{ij})}^{m}{||x_{j}% -v_{i}||}^{2}}}{N_{i}\sum^{c}_{k=1}{{||v_{k}-v_{i}||}^{2}}}}$ (24)

where $N_{i}$ is the number of samples in the $i^{\text{th}}$ cluster. $S C$ is useful for comparing different partitions having an equal number of clusters. A lower value of $S C$ indicates a better partition. Separation index (S): Unlike the partition index (S), the separation index uses a minimum distance separation for partition validity [67].

$\displaystyle S(c)=\frac{\sum^{c}_{i=1}{\sum^{N}_{j=1}{{(u_{ij})}^{2}{||x_{j}-% v_{i}||}^{2}}}}{N{\min}_{i,k}{||v_{k}-v_{i}||}^{2}}$ (25)

Xie and Beni’s Index (XB): This is one of the most widely used fuzzy clustering validity indices proposed by Xie and Beni in 1991 with parameter $m$ set to 2. It aims to quantify the ratio of total variation within clusters and cluster separation. This index was developed as a cluster validity ratio function measuring the fuzzy cluster compactness through the distance of samples in a cluster from their cluster centers and the cluster separation represented by the distance between cluster centers [66]. Considering a fuzzy partition of the data set $X=\{x_{j};j=1,\ldots,n\}$ with $v_{i}(i=1,\ldots,n_{c}\}$ the centers of each cluster and $u_{ij}$ the membership of data point $j$ to cluster $i$ , the index $X B$ can be calculated as follows:

$\displaystyle XB(c)=\frac{\sum^{c}_{i=1}{\sum^{N}_{j=1}{{(u_{ij})}^{m}}{||x_{j% }-v_{i}||}^{2}}}{N{\min}_{i,j}{||x_{j}-v_{i}||}^{2}}$ (26)

A smaller value of the $X B$ index indicates a better quality of clustering, meaning more compact and well separated clusters. However, the $X B$ index decreases monotonically as $c$ approaches $n$ (number of clusters $n_{c}$ becomes very large and close to $n$ ). Dunn index: The Dunn index ( $D I$ ) was introduced by J.C. Dunn in 1974 as a metric for evaluating clustering algorithms. It defines the ratio between the minimum intra-cluster distance and the maximum inter-cluster distance in order to identify the compact clusters sets, where the means of the different clusters are sufficiently distant, from the cluster within the variance [66]. The $D I$ index is given by the following equation:

$\displaystyle{DI}_{c}={\min}_{i\in c}\left\{{\min}_{j\in c,i\neq ≠j}\left\{% \frac{{{d(C_{i},C_{j})}}}{{\max}_{k\in C}\textit{diam}(C_{k})}\right\}\right\}% ={\min}_{i\in c}\left\{{\min}_{j\in c,i\neq j}\left\{\frac{{\min}_{x\in C_{i},% y\in C_{j}}d(x,y)}{{\max}_{k\in C}\{{\max}_{x,y\in C}d(x,y\}}\right\}\right\}$ (27)

where ${d(C_{i},C_{j})}$ is a function of dissimilarity between two clusters $C_{i}$ and $C_{j}$ defined by ${d(C_{i},C_{j})}={\min}_{x\in C_{i},y\in C_{j}}d(x,y)$ which denotes the minimum distance between two elements $x_{i}$ and $y_{j}$ (objects of different clusters) classified separately, and $\textit{diam}(C_{k})$ the diameter of a cluster, which can be considered as a measure of cluster dispersion and defined as $\textit{diam}(C)=\begin{array}[]{c}\max\\ x,y\in C\\ \end{array}d(x,y)$ where $\begin{array}[]{c}\max\\ x,y\in C\\ \end{array}d(x,y)$ denotes the maximum distance between two elements classified together (objects of the same cluster). The Dunn index is limited to the interval [0, 1] and should be maximized to indicate better clustering (the presence of compact and well separated clusters). Davies-Bouldin index: The Davies-Bouldin (DB) index is a measure of similarity between clusters that is defined from a cluster dispersion measure and a dissimilarity measure between two clusters. This index treats each cluster individually and seeks to measure the average similarity between each cluster $C_{i}$ , ( $i=1,..N$ ), and the cluster closest to it (similar) [68]. It is the average of the maximum ratio between the distance of a point in the center of its cluster and the distance between two cluster centers, and is formulated as follows:

$\displaystyle\textit{DBI}=\frac{{1}}{N}\sum^{n}_{i={1},i\neq j}{\max}\left(% \frac{\sigma_{i}+\sigma_{j}}{d(c_{i},c_{j})}\right)$ (28)

where $N$ is the number of clusters, $\sigma_{i}$ is the average distance of all objects in cluster $i$ to their cluster center $c_{i}$ , $\sigma_{j}$ is the average distance of all objects in cluster $j$ to their cluster center $c_{j}$ , and $d(c_{i},c_{j})$ is the distance from the cluster centers $c_{i}$ and $c_{j}$ . The small DBI values correspond to compact clusters whose centers are far from each other. The best partition is therefore the one that minimizes the average of the calculated value for each class. In other words, the best partition is the one that minimizes the similarity between classes.

4.2.2 Evaluation criteria for service availability and system performance

During each cycle, measurements are captured and analyzed at the end of the simulation to estimate the performance of the proposed system in terms of service availability and meet the quality of service required by users. We have focused on the following performance indicators:

Success Search, the search for a requested service takes on a positive value if the Partial Cloud Server finds at least one participant that hosts the requested service. Failure Search, the search for a requested service fails if the Partial Cloud Servers cannot find a service provider. Success Result, expresses the success of result returned by the participants after execution of the requested services. Failure Result, which means the failure of the participating nodes to correctly provide the result of the requested service. Number of Resource Nodes, is the number of nodes that host the requested service. Maximum Average Service Availability is the average of the maximum probability of the services being available at the time of the request and being able to perform their required tasks correctly to the users. Minimum Average Service Availability is the minimum average probability that services will be available at the time of the request and able to perform their required tasks correctly to users.

4.3 Results and discussion

We conducted 500 tests for each scenario in order to assess the impact of the proposed approach on improving services and entities availability and system performance. The experimental results are expressed as the mean variance of all identical experiments under the scenarios described above, and varying under five other simulation scenarios depending on the size of the P2P network (size: 10000, 20000, 30000, 40000, 50000).

In order to conduct comparative experiments, we first examined our proposed Hybrid Fuzzy Clustering Method “HAC-TFCM” in scenarios of clustering of multidimensional user profile data by comparing it with widely used methods, namely the Fuzzy C-Means (FCM) and Hierarchical Ascending Classification (HAC) methods. Experimental data are dynamic in nature and are collected during the network’s operation within each experiment. Moreover, these data are considered as initial inputs for all the algorithms and formulas used in this study. In the case of FCM (Fuzzy C-Means), we performed a series of clustering validation experiments that reflect the properties of the environment by varying the number of clusters between 2 and 6. The U membership matrix was randomly initialized. The value of the fuzzification parameter $m$ was set to 2, $\varepsilon=$ 0.01 with maximum number of iterations $=$ 100 in all simulation experiments.

In this section, the proposed hybrid clustering approach has been evaluated with six previously introduced internal validity indices. The diagram in Fig. 6 shows the trend in the average variation of various cluster validity indices over the experimental dataset that are captured during the simulation scenarios. To facilitate visualization, the minimum value of all indicators is 0 and the maximum value is 1 per normalization treatment. Note that the clustering effect is best when all indicators are as small as possible except the PC (Partition Coefficient) index.

Figure 6.

Average internal validity index values of hybrid fuzzy clustering mechanism, fuzzy C-means and hierarchical ascending classification.

Table 4

Validation of fuzzy C-means, hierarchical ascending classification and hybrid hierarchical ascending classification with thresholded fuzzy C-means, applied on client profile dataset

$m=$ 2, $\epsilon$ 0.01		FCM					HAC	Hybrid HAC-TFCM
Network size	Index	Cluster
		2	3	4	5	6
10000 nodes	SC	0.4498	0.3918	0. 411	0. 407	0.418	0.466	0.391
	PC	0.6047	0.6867	0.6446	0.6241	0.8234	0.7387	0.856
	PE	0.192	0.321	0.294	0.337	0.369	0.426	0.196
	XB	0.156	0.181	0.231	0.310	0.242	0. 36	0.115
	DI	0.146	0.1497	0.1588	0.1807	0.175	0.1696	0.193
	DB	0.402	0.421	0.386	0.364	0.351	0.477	0.314
20000 nodes	SC	0.475	0.431	0.421	0.415	0.345	0.391	0.312
	PC	0.642	0.597	0.601	0.776	0.868	0.632	0.871
	PE	0.215	0.226	0.237	0.249	0.304	0.515	0.207
	XB	0.172	0.167	0.184	0.164	0.106	0.197	0.091
	DI	0.143	0.171	0.112	0.103	0.154	0.285	0.185
	DB	0.412	0.386	0. 389	0. 377	0. 364	0.426	0.291
30000 nodes	SC	0.510	0.485	0.471	0.382	0.341	0.394	0.314
	PC	0.590	0.633	0.670	0.669	0.792	0.617	0.837
	PE	0.215	0.229	0.336	0.326	0.364	0. 453	0.224
	XB	0.295	0.326	0.314	0.354	0.307	0.294	0.184
	DI	0.217	0.196	0.222	0. 192	0. 204	0.325	0.212
	DB	0.399	0.394	0.386	0.412	0.371	0.501	0.361
40000 nodes	SC	0.547	0.612	0.584	0.513	0.490	0.435	0.425
	PC	0.623	0.601	0.671	0.741	0.707	0.694	0.893
	PE	0.206	0.216	0.258	0.312	0.359	0.610	0.176
	XB	0.312	0.298	0.241	0.261	0.208	0.265	0.205
	DI	0. 157	0. 132	0.159	0.211	0.193	0.265	0.194
	DB	0.421	0.429	0.405	0.381	0.364	0.396	0.350
50000 nodes	SC	0.561	0.430	0.429	0.436	0.394	0.424	0.243
	PC	0.683	0.751	0.813	0.816	0.859	0.787	0.871
	PE	0.264	0.324	0.339	0.321	0.354	0.494	0.189
	XB	0.348	0.375	0.3097	0.298	0.266	0.237	0.196
	DI	0.243	0.177	0.172	0.152	0.154	0.312	0.223
	DB	0.412	0.433	0.397	0.384	0.371	0.425	0.307

The results show that the $P C$ (Partition Coefficient) index values vary in a range above 0.5 and are close to $+$ 1 for all algorithms. While, the proposed clustering mechanism can reach $P C$ (Partition Coefficient) values higher than those noted with the FCM (Fuzzy C-Means) and HAC (Hierarchical Ascending Classification) algorithm with a mean of 0.87 indicating good clustering and also a strong underlying structure. In contrast to the $P C$ (Partition Coefficient) index, smaller $P E$ (Partition Entropy) and $S C$ (Partition Index) values were found by the FCM (Fuzzy C-Means) and HAC-TFCM (Hybrid Hierarchical Ascending Classification with Thresholded Fuzzy C-Means) ( ${SC}_{\textit{HAC-TFCM}}=$ 0.335, ${SC}_{\textit{FCM}}=$ 0.454, ${PE}_{\textit{HAC-TFCM}}=$ 0.198, ${PE}_{\textit{FCM}}=$ 0.286) algorithm showing the presence of well-separated clusters, while the $P E$ (Partition Entropy) values for HAC (Hierarchical Ascending Classification) approach the upper limit of the range indicating the inability of the algorithm to extract any clustering structure from the data set. In addition, the minimum values of the $X B$ (Xie and Beni’s) and $D B$ (Davies-Bouldin) validity measures by FCM (Fuzzy C-Means) indicated a better quality of clustering when there were 6 classes in the data. Concurrently, the HAC-TFCM (Hybrid Hierarchical Ascending Classification with Thresholded Fuzzy C-Means) algorithm can achieve minimum values of $X B$ (Xie and Beni’s) and $D B$ (Davies-Bouldin) even better than FCM (Fuzzy C-Means) and HAC (Hierarchical Ascending Classification). As a result, the proposed hybrid algorithm is able to automatically detect the ideal number of clusters for a better final partitioning of the data set. In addition, it should be noted that the Dunn index shows that the FCM (Fuzzy C-Means) classifier with 5 clusters provides the best results, in contrast to other indices that designate HAC-TFCM (Hybrid Hierarchical Ascending Classification with Thresholded Fuzzy C-Means) as more accurate. The experimental results verify the superiority of the hybrid HAC-TFCM (Hybrid Hierarchical Ascending Classification with Thresholded Fuzzy C-Means) algorithm in efficiently and reliably clustering the client profile data set, creating better separated and more meaningful clusters with high compactness. However, the HAC (Hierarchical Ascending Classification) algorithm did not result in well separated and dense clusters.

The main objectives of the user profile clustering experiences were to study the ability of fuzzy clustering to discover good client relationships and to design a highly available P2P based cloud SaaS service system to improve the client experience. Now, we will show the impact of the proposed approach on the variation of Success_Search and Failure_Search percentages per cycle for the different simulation scenarios, Fig. 7.

Figure 7.

Success search percentage vs failure search percentage.

According to the statistics, we initially observe that during the first simulation cycles, the percentage of Failure_Search is greater than the percentage of Success_Search. From the 4 ${}^{\text{th}}$ simulation cycle, the percentage of Success_Search starts progressing surpassing the percentage of Failure_Search in all system scenarios and during the rest of the stimulation. In addition, the results show that the percentage of Success_Search obtained by the system implementing the proposed approach increases constantly over time and begins to stabilize from the 16 ${}^{\text{th}}$ cycle reaching a maximum of 88.19% at the 20 ${}^{\text{th}}$ cycle. According to these statistics, we can see that the Success_Search has progressed continuously in the system adopting FCM (Fuzzy C-Means) reaching a maximum percentage of 80.1% in the 24 ${}^{\text{th}}$ cycle as well as in the system adopting HAC (Hierarchical Ascending Classification) reaching a maximum of 71.18% in the 24 ${}^{\text{th}}$ cycle and then stagnating until the end of the simulation. Furthermore, a simultaneous similar variation can be observed in the percentage of Success_Search in the system proposed in [6], reaching a maximum of 75.26% in the 23 ${}^{\text{rd}}$ cycle. This is done by having network peers that host gradually the services and become participants in the system. In contrast, the percentage of Failure_Search in both systems increases at the beginning of the simulation because at this stage the system is still in its early stages of execution and there are fewer participants. From the 3 ${}^{\text{rd}}$ cycle, we can observe a decreasing trend in the percentage of Failure_Search over time that declines continuously and tends towards values lower than 30% at the end of the simulation where the system implementing the proposed approach achieves values that tend towards 10%. However, we can clearly see that the percentage of Success_Search has increased significantly over time in the system implementing the proposed approach reaching a peak of 60% in the 4 ${}^{\text{th}}$ cycle which is an obvious consequence since the proposed approach adopts a clustering mechanism that virtually groups together peer service providers with similar profiles in specific regions which will facilitate their detection within reasonable time intervals. The search process only uses suppliers belonging to the virtual group that reflects the best similarities of the peer requester, thus allowing reduction of search space and improvement of response time. Consequently, the proposed approach improved the percentage of Success_Search by an average of 10% during the simulation scenarios compared to the system proposed in [6]. Moreover, the Success_Search results obtained by the system adopting the proposed approach surpass those obtained by the system using FCM (Fuzzy C-Means) and HAC (Hierarchical Ascending Classification), offering an improvement of 8% and 17% successively. Note that the results of the system proposed in [6] are better than that obtained by the system implementing the HAC (Hierarchical Ascending Classification) algorithm. As a consequence, an assumption can be made that the quality of clustering affects system performances so that when the final partitioning of client profiles data is optimal, the search process is more efficient.

In service-oriented systems, the major concern is to ensure the successful provision of services that reflect their ability to provide the resource provisioning function on demand. Figure 8 shows the evolution of the percentage of the two measures: Success_Result and Failure_Result according to the demands made in the different simulation scenarios. These statistics are represented by assuming the same failure rate for service providers (0.25%) as well as the probability of no failure in peer providers in different systems.

Figure 8.

Success result percentage vs failure result percentage.

According to these statistics, it can be observed that the percentage of Success_Result in the system proposed in [6] has continuously increased to reach a maximum of 78.91% in the 19 ${}^{\text{th}}$ cycle and then it stagnates until the end of the simulation. This change is associated with a variation in the Failure_Result, which increases over time until the 8 ${}^{\text{th}}$ cycle and then gradually decreases and stabilizes until the end of the simulation reaching a minimum of 25.47%. In parallel, a significant increase in the percentage of Success_Result over the first six cycles by the system implementing the proposed approach reaching a rate of 68.5%. This variation is associated with a rapid increase in the percentage of Failure_Result reaching around 40.32%. From the sixth cycle, the percentage of Success_Result continues to increase rapidly reaching a maximum value of 86.65% in the 15 ${}^{\text{th}}$ cycle and then begins to stabilize until the end of the simulation. While the percentage of failure results begin to decrease over time and are then associated with a stagnation that tends towards a value of 13%. Besides, we can see that the percentage of Success_Result and Failure_Result in the systems implementing FCM (Fuzzy C-Means) and HAC (Hierarchical Ascending Classification) follows the same variation trend knowing that the percentage of Success_Result reaches values lower on average of 6% and 12.3% successively than those obtained by the system adopting our proposed approach. While it achieves a maximum of 82.2% and 75.41% in the 19 ${}^{\text{th}}$ and 22 ${}^{\text{nd}}$ cycle successively before it stagnates until the end of the simulation. In contrast, the percentage of Failure_Result in these systems attained average values of 5% and 7% higher than those obtained by the system adopting our proposed approach where this criterion increases in the first simulation cycles to 53% and 48% before it begins to decline and stabilize towards percentages of 20% and 25%. As a result, the improvement in Success_Result and Failure_Result rates indicates superior reliability and performance of the proposed approach on the SaaS provisioning system compared to the HAC (Hierarchical Ascending Classification) and FCM (Fuzzy C-Means) clustering algorithms. In addition, it reveals from these findings that the rate of efficient service provision in the proposed architecture is significantly improved on average 86.91% compared to 78.91% obtained by the system proposed in [6]. As 1% success is more important, the increase in the system implementing the proposed approach is 9.69% on average compared to the one that does not use the clustering mechanism. The latter generates a virtual layer in which similar peer providers with different and complementary functional features can be grouped together. By maintaining diversity, the system can maintain continuous operability by peers who may belong to different and similar geographic regions and time zones. In other words, this allows services to be hosted and executed by peers with varying properties in different geographic locations and to be present in multiple physical areas. This will have a positive impact on the availability of required services and system performances.

Here, we will study the behavior of the probability of SaaS service availability taking into account the impact of the cloud-P2P service provisioning system conception.

Figure 9.

Mean service availability percentage.

Figure 9 shows the variation in the percentage of service availability in the different simulation scenarios. The availability of services in the system implementing the proposed approach increases rapidly at the beginning of the simulation, reaching a percentage of 61.38% at the 5 ${}^{\text{th}}$ cycle. However, the availability decelerates between the 5 ${}^{\text{th}}$ and 11 ${}^{\text{th}}$ cycles and then slightly fluctuates between the 11 ${}^{\text{th}}$ and 19 ${}^{\text{th}}$ cycles as the system reorganizes their virtual infrastructure periodically. This variation does not decline to 80% until it stabilizes, reaching a maximum availability of 89.19% at the 21 ${}^{\text{th}}$ cycle. In addition, the service availability percentages in systems implementing clustering algorithms including FCM (Fuzzy C-Means) and HAC (Hierarchical Ascending Classification) vary similarly to the previous system reaching a maximum availability of 81.27% and 76.12% successively. However, the system implementing the proposed approach has succeeded to achieve average improvement values of 9% and 14% compared to the system adopting the FCM (Fuzzy C-Means) and HAC (Hierarchical Ascending Classification) clustering algorithm successively. Furthermore, Fig. 9 also illustrates that the service availability in the system proposed in [6] increases slowly at the beginning of the simulation reaching a peak of 48.4% in the 7 ${}^{\text{th}}$ cycle. In addition, we note that the service availability deteriorates considerably in several instants between the 12 ${}^{\text{th}}$ and 21 ${}^{\text{th}}$ cycles and then starts to stabilize over time reaching a maximum of 78.14% in the 24 ${}^{\text{th}}$ cycle. The volatile behavior of the availability percentage is due, in our opinion, to the fact that the service availability depends entirely on the individual availability of the peers performing on the system, which means that all their associated services may be affected by their own failures.

As illustrated in Fig. 9, the percentage variation of service availability in the system adopting the proposed approach follows a similar way as other systems, despite the fact that the service availability percentage in the system adopting the proposed approach remains higher [6]. Moreover, by comparing the three curves associated to the scenarios in which one of the clustering algorithms is used, we notice that as the algorithm has lower performance on the experimental data set and the quality of clustering decreases, the improvement rate of service availability declines and gradually approaches the system of [6] where clustering is not used.

Table 5

Average, min and max values of the curves of Fig. 10

	Network size	Average availability	Max availability	Min availability
	(nodes)	(%)	(%)	(%)
System implementing the proposed approach	10000	87.18	87.97	63.51
	20000	87.67	89.25	65.46
	30000	89.05	90.34	70.14
	40000	88.24	88.85	68.83
	50000	88.56	89.54	69.25
System implementing FCM	10000	78.05	83.21	59.28
	20000	78.34	80.92	56.75
	30000	78.23	82.18	59.07
	40000	80.15	80.36	56.04
	50000	79.36	79.68	54.61
System implementing HAC	10000	73.23	75.31	29.16
	20000	73.05	75.20	27.88
	30000	75.15	75.42	31.28
	40000	74.36	77.03	35.94
	50000	74.34	77.64	36.47
System proposed in [6]	10000	75.17	77.31	39.22
	20000	74.61	77.22	38.91
	30000	75.82	77.65	40.36
	40000	78.54	80.03	41.56
	50000	76.09	78.52	40.62

The results of these simulations are summarized in Table 5 in which the statistics show an incremental trend between clustering quality and service availability, which again justifies the assumption that an optimal clustering model is necessary for the design of a high-performance and reliable service provisioning system. In addition, the statistics presented show that the clustering and virtual layer formation method considered by the proposed approach has a considerable effect in improving the percentage of SaaS service availability in a P2P-based cloud architecture compared to the architecture proposed in [6]. Efficient provider distribution involves the formation of condensed zones of required services hosted or executed by a set of peers. As a result, the condensed location of a required service in a specific region of the virtual layer increases the probability of its availability at the time of its request.

The proposed approach can achieve an Average Service Availability of 88% versus 76% with a total increase of 12% compared to the system proposed in [6]. Moreover, the system implementing the proposed approach can reach a Maximum Average Service Availability of 89.19% and Minimum Average Service Availability of 67.43% in the availability calculations versus 78.94% and 40.13 % in the architecture proposed by [6], which means that even in the failure cases the system, remains relatively efficient. Therefore, as there are more peer providers available belonging to one or more clusters represented by virtual nodes in addition to the communication links, the system is more available and responds faster. Besides, it seems that our approach gives a perfect combination of calculations using time-based availabilities, for better numerical analysis, and activity-based availabilities, for more accurate information about system during its operation, and presence-based availabilities monitoring the overall state of the system entities.

The proposed approach responds widely to the large-scale nature allowing the conception of a SaaS service provision system based on a P2P network able to find the best compromise between client satisfaction in terms of quality of service and the cost of the system.

5. Comparative discussion

Here the most relevant existing works for improving availability in the cloud are compared with our proposed work in order to demonstrate that the proposed work is better one. It should be noticed that these works are contrasted with our proposed approach in their performance analysis methods where each of them adopts a distinct experimental environment and is based on evaluation criteria different from those used in our paper. Therefore, the comparison scope will be done in terms of the proposed techniques’ conception as well as their impact on the cloud environment’s performance and especially on the availability of services and resources.

In order to ensure high availability and reliability, cloud service providers introduce various mechanisms such as redundancy, scheduling and fault tolerance in their cloud systems. Following these mechanisms, the authors in [14, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 40, 41, 42] have presented various effective strategies. Although these proposed approaches have contributed to enhance data availability and accessibility, prevent network disruptions and ensure system continuity. They still suffer from problems related to their core operation, which relies mainly on the need for global knowledge of networks and their complex exploitations stored and computed during their processing. In addition to the volume of dynamic data analyzed in large-scale distributed environments, these techniques typically produce huge amounts of data flowing through the network, resulting an unnecessary increase in effort and cost in terms of: storage, processor, financial overhead, maintenance overhead, complex and consistent data management. Moreover, the processing and execution delays in large networks, the increase in peer inactivity time and the limitation of network bandwidth create bottlenecks that are very sensitive to concurrent data access in the distributed environment.

In contrast to these works, the proposed approach meets the requirements of low-cost availability by maintaining a virtual and dynamic layer added to P2P network based-SaaS cloud service system architecture. This layer is built using a hybrid clustering mechanism of client profile vectors that are dispatched, according to their profile’s similarities, in several clusters. It enables the distribution of peer providers and the creation of condensed areas of service hosted on multiple nodes, thus enhancing the availability probability degree of services in specific regions at the time of demand. In addition, the service search process is initially carried out in the virtual layer instead of using all system users and in case of failure it will be launched in the real “Partial Cloud Servers” layer. The proposed approach provides high availability of the services as well as the peers holding them, efficient communication and network efficiency. The proposed approach ensures high availability of the services as well as the peers holding them, efficient communication and network efficiency. In other words, our system is able to ensure that clients interact continuously and correctly with SaaS applications and services, and that peer churn (normally by disconnection, abnormally by failure) does not affect system performance. Therefore, the proposed approach provides an optimal fault-tolerant infrastructure that responds largely to the dynamic nature of peers as well as their varying behaviors in the real network layer of the service delivery system. It allows to reduce the search space and communication cost, maintain low network traffic, improve the response time and success rate of service execution.

By adopting various clustering methods, other authors [48, 49, 50, 51, 52, 53, 54] have proposed several techniques in order to maintain adequate reliability and availability to their cloud users’ needs. However, the proposed techniques suffer from several challenges including: high network traffic, large overhead, low load balancing, reduced efficiency in large scale environments, simple data handling and clustering methods. As a clustering mechanism, the proposed approach used a fuzzy hybrid clustering method “HAC-TFCM” to perform a new clustering of network peers besides their clustering in the real network layer. The proposed clustering technique surpasses the techniques presented in [48, 49], as it allows each client to be automatically associated with the appropriate number of clusters judged to be close without a priori knowledge of the cluster number and how many clusters it should belong to. Compared to [51], our proposed approach was able to manage complex and dynamic data consisting of profile vectors of the system’s clients which determines their own network behavior. As compared to [52, 54], our system is scalable and its performance remains stable in large-scale environments allowing the conception of a P2P network-based SaaS service delivery system able to achieve the best trade-off between client satisfaction in terms of service quality and system cost. In addition, our paper is distinct from the works discussed below as it offers a service availability measurement model that considers different entities at several levels of the system and uses various types of o availability.

6. Conclusion

In the contemporary technological world, technology requires an effective approach to minimize overall costs and improve the overall reliability of the system by offering the best quality of services. In this paper, we proposed an approach to improve the persistent availability of SaaS services in a “Hybrid P2P Network-Based Remote Treatment SaaS cloud computing” environment. This is made in order to meet the requirements of on-demand service provisioning and to maintain cloud computing performances by overcoming the problems of system failures and peer dynamics. This approach primarily tracks the dynamic and rapid rhythm of these peers by analyzing their individual behaviors and representing them in the form of a matrix of client profile vectors. Besides, a Hybrid Fuzzy Clustering Mechanism has been proposed by combining in a sequential hybridization process the “Hierarchical Ascending Classification” algorithm and the “Thresholded Fuzzy C-Means”, allowing each of the client’s profile vector to be automatically associated to the appropriate number of clusters considered close without any prior knowledge of the cluster number and the number of clusters to which it should belong. It aims to provide a virtual and optimal infrastructure to organize the system peers based on their profile similarity into distinct clusters where each cluster is represented by a virtual node forming together a virtual layer. This layer provides both the distribution of peer service providers and the formation of dense zones of each service that interests a set of peers who host or use them, which greatly improves the probability of availability of these services in specific regions of the virtual layer at the request time. In addition, a service availability measurement model has been proposed, which is mainly based on the use of the virtual layer of the system and takes into account different entities at different levels of the system including the probability of availability of service providers and the probability of availability of virtual nodes. To obtain the best performance in a service provision system, we used different types of availability existing in the literature to measure the availability of these entities: “time-based availability, activity-based availability, availability based on the presence of k-of-n” which are refined to include special considerations in their calculations due to their particular use.

We examined the proposed hybrid fuzzy clustering method based on six internal validity indices and the experimental results verify the superiority of the hybrid fuzzy clustering algorithm “HAC-TFCM” compared to FCM and HAC in the efficient and reliable clustering of client profile data set creating better separated and more meaningful clusters with high compactness. Furthermore, the results of the experimental evaluation with PeerSim provide convincing evidence that the proposed approach can effectively improve the probability of SaaS service availability and the reliability of the cloud-P2P system for remote service provisioning offering an average service availability of 89% with a 7% improvement. Consequently, the proposed approach provides irrefutable proof that it largely responds to the large-scale nature of distributed systems and is able to achieve the best compromise for maintaining quality of service in terms of availability, performance and cost.

As future perspectives, it would be interesting to extend the proposed approach by taking into account other measures of cloud services quality and to test our approach on benchmark datasets.

Footnotes

Author’s Bios

	Abir Achache Phd student at the University of Badji Mokhtar Annaba – UBMA – Algeria. Miss Achache studied at UBMA where she obtained a diploma in “Computer Science” and a Master’s degree in “Information and Communication Science and Technology”. Her research interests include Artificial Intelligence, parallel and distributed computing, P2P networks and Cloud Computing. She contributes to the development and improvement of the Cloud Computing architecture by providing her knowledge and experience to the benefit of students and professionals.
	Abdelhalim Baaziz He received his MsC degree in Computer Science from Badji Mokhtar Annaba University, Algeria, in 1994 and Ph.D degree in software engineering from Ecole poly technique Montreal University, Canada, in 2008. He was the head of the Department of Computer Science at Badji Mokhtar Annaba University from 2013 to 2016. His research interests include P2P and Adhoc networks, Cloud Computing and software sharing.
	Toufik Sari He received his MsC. and Ph.D. degrees in Computer Science from Badji Mokhtar Annaba University in 2000 and 2007, respectively. His research interests include: Artificial Intelligence, Signal, Image and Video Processing, Speech and language Processing, Feature Extraction, Computational Intelligence, Document Analysis, Information Retrieval, Computer Communications and Networks.

References

Zhang

Cheng

and Boutaba

, Cloud computing: State-of-the-art and research challenges, Journal of Internet Services and Applications 1 (2010), 7–18.

Syed

H.I.

and Baig

N.A.

, Survey on cloud computing, International Journal of Emerging Technology and Advanced Engineering 3 (2013), 308–312.

Sandanayake

T.C.

and Jayangani

P.G.C.

, Current trends in software as a service (SaaS), International Journal for Innovation Education and Research 6 (2018), 221–234.

Kaur

, Software as a service: A brief study, International Research Journal of Engineering and Technology 2 (2015), 789–792.

and Han

, Convergence P2P cloud computing, Peer-to-Peer Networking and Applications 11 (2018), 1153–1155.

Achache

Baaziz

and Sari

, A Hybrid P2P Network-Based Remote Treatment SaaS Cloud Computing, in: 5𝑡ℎ International Conference on Green Computing and Engineering Technologies (ICGCET), Casablanca, Morocco, September 17–19, 2019.

Maurya

R.K.

Pandey

and Kumar

, A survey of peer-to-peer networks, International Journal of Advanced Research in Computer and Communication Engineering 5 (2016), 292–298.

Steinmetz

and Wehrle

, Peer-to-Peer Systems and Applications, 1

{}^{\text{st}}

Edition, Springer, Berlin, Heidelberg, Germany, 2005.

Da-Costa

Marchand

Richard

and Vincent

J-M.

, Resources availability for Peer to Peer systems, in: 20𝑡ℎ International Conference on Advanced Information Networking and Applications, Washington, United States, 2006, pp. 6–28.

10.

Dhote

R.A.

and Deshpande

S.P.

, Data mining with cloud computing: An overview, International Journal of Advanced Research in Computer Engineering & Technology 5 (2016), 211–214.

11.

Ahuja

S.P.

and Mani

, Availability of services in the era of cloud computing, Network and Communication Technologies 1 (2012), 2–6.

12.

Karimi

O.B.

Yousefi

Fathy

and Mazoochi

, Availability in Peer to Peer Management Networks, Challenges for Next Generation Network Operations and Service Management Conference, in: 11𝑡ℎ Asia-Pacific Network Operations and Management Symposium, Beijing, China, October 22–24, 2008, pp. 552–555.

13.

Pavlik

Sobeslav

and Horalek

, Statistics and Analysis of Service Availability in Cloud Computing, in: 18th International Database Engineering and Applications Symposium, Porto Portugal, July 7–9, 2014, pp. 310–313.

14.

Wei

Veeravalli

Gong

Zeng

and Feng

, CDRM: A Cost-Effective Dynamic Replication Management Scheme for Cloud Storage Cluster, in: 2010 IEEE International Conference on Cluster Computing, Heraklion, Greece, September 20–24, 2010, pp. 188–196.

15.

Bandela

Gadde

and Pabboju

, Survey on cloud computing technologies and security threats, International Journal of Engineering Research and Technology 2 (2015), 296–308.

16.

Vani

Cynthia

and Priya

, Availability in cloud computing, International Journal of Innovative Research in Information Security 4 (2015), 11–15.

17.

Mukherjee

Shaw

Haldar

and Changdar

, A survey of data mining applications and techniques, International Journal of Computer Science and Information Technologies 6 (2015), 4663–4666.

18.

Hussain

, Survey on current trends and techniques of data mining research, London Journal of Research in Computer Science and Technology 17 (2017), 7–16.

19.

Kamel

M.S.

and Selim

S.Z.

, A thresholded fuzzy c-means algorithm for semi-fuzzy clustering, Pattern Recognition 24 (1991), 825–833.

20.

Sasirekha

and Baby

, Agglomerative hierarchical clustering algorithm – a review, International Journal of Scientific and Research Publications 3 (2013), 2250–3153.

21.

Schmitt

and Steinmetz

, The Effectiveness of Realistic Replication Strategies on Quality of Availability for Peer-to-Peer Systems, in: Third International Conference on Peer-to-Peer Computing, Sweden, September 1–3, 2003, pp. 57–64.

22.

Netes

, End-to-End Availability of Cloud Services, in: The 22𝑛𝑑 Conference of Open Innovations of Association, Jyvaskyla, Finland, May 15–18, 2018, pp. 198–203.

23.

Dunn

R.J.

Zahorjan

Gribble

S.D.

and Levy

H.M.

, Presence-based availability and P2P systems, in: 5𝑡ℎ IEEE International Conference on Peer-to-Peer Computing, Konstanz, August 31–September 2, 2005, pp. 209–216.

24.

Karimi

O.B.

Yousefi

Fathy

and Mazoochi

, Availability measurement in Peer to Peer Network Management Systems, in: Third IEEE International Conference on Digital Information Management, (ICDIM), London, UK, November 13–16, 2008, pp. 745–750.

25.

Nabi

Toeroe

and Khendek

, Availability in the cloud: State of the art, Journal of Network and Computer Applications 60 (2016), 54–67.

26.

Grover

, A study of various fuzzy clustering algorithms, International Journal of Engineering Research 3 (2014), 177–181.

27.

Saadat

and Rahmani

, PDDRA: A new pre-fetching based dynamic data replication algorithm in data grids, Future Generation Computer Systems 28 (2012), 666–681.

28.

Sashi

and Thanamani

A.S.

, Dynamic replica management for data grid, International Journal of Engineering and Technology 2 (2010), 329–333.

29.

Rama

and Reddy

, Data replication system in cloud based on data mining techniques, International Journal of Advanced Research in Computer and Communication Engineering 2 (2013), 4216–4221.

30.

Hussein

and Mousa

, A light-weight data replication for cloud data centers environment, International Journal of Innovative Research in Computer and Communication Engineering 2 (2014), 2392–2400.

31.

Rahmati

and Rahmani

A.M.

, Data replication-based scheduling in cloud computing environment, Journal of Advances in Computer Engineering and Technology 3 (2017), 75–80.

32.

Sun

D.W.

Chang

G.R.

Gao

Jin

L.Z.

and Wang

X.W.

, Modeling a dynamic data replication strategy to increase system availability in cloud computing environments, Journal of Computer Science and Technology 27 (2012), 256–272.

33.

Bsoul

Al-Khasawneh

Kilani

and Obeidat

, A threshold-based dynamic data replication strategy, The Journal of Supercomputing 60 (2012), 301–310.

34.

Park

Kim

and Yoon

, Dynamic Data Grid Replication Strategy Based on Internet Hierarchy, in: International Conference on Grid and Cooperative Computing, Wuhan, China, October 21–24, 2003, pp. 383–846.

35.

Sashi

and Thanamani

A.S.

, Dynamic replication in a data grid using a modified BHR region based algorithm, Future Generation Computer Systems 27 (2011), 202–210.

36.

Zhou

and Yan

, A Hierarchical Data Replication Method in ad hoc networks, in: 2𝑛𝑑 International Conference on Computer Engineering and Technology, Bali Island, Indonesia, Mars 26–29, 2010, pp. 23–27.

37.

Vijaya

C.K.

and Ramachandhra

G.A.

, Optimization of large data in cloud computing using replication methods, International Journal of Computer Science and Information Technologies 5 (2014), 3034–3038.

38.

Noraziah

Fairuzullah

Azila

C.F.

Azuan

Aqtar

Azlina

and Herawan

, Fault Tolerance On Binary Vote Assignment Cloud Quorum Replication Technique, in: International Conference on Computer Science and Information Technology, Yogyakarta, Indonesia, June 16–18, 2013, pp. 2–6.

39.

Myint

and Naing

, Management of data replication for PC cluster based cloud storage system, International Journal on Cloud Computing: Services and Architecture 1 (2011), 31–41.

40.

Thein

and Park

J.S.

, Availability analysis of application servers using software rejuvenation and virtualization, Journal of Computer Science and Technology 24 (2009), 339–346.

41.

Singh

and Chhabra

, High Availability of Clouds: Failover Strategies for Cloud Computing using Integrated Checkpointing Algorithms, in: International Conference on Communication Systems and Network Technologies, Rajkot, India, May 11–13, 2012, pp. 698–703.

42.

Jhawar

and Piuri

, Adaptive Resource Management for Balancing Availability and Performance in Cloud Computing, in: The 10𝑡ℎ International Conference on Security and Cryptography, Reykjavik, Iceland, July 29–31, 2013, pp. 254–264.

43.

Liu

Boehm

and Alis

, Change Detection of Mobile Lidar Data Using Cloud Computing, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences, Prague, Czech Republic, July 12–19, 2016, 309–313.

44.

Zhang

Zhani

M.F.

Jabri

and Boutaba

, Venice: Reliable Virtual Data Center Embedding in Clouds, in: IEEE Conference on Computer Communications, Toronto, Canada, April 27–Mai 2, 2014, pp. 289–297.

45.

Makhsous

S.H.

Gulenko

Kao

and Liu

, High Available Deployment of Cloud-Based Virtualized Network Functions, in: International Conference on High Performance Computing and Simulation (HPCS), Innsbruck, Austria, July 18–22, 2016, pp. 468–475.

46.

Benz

and Bohnert

, Dependability Modeling Framework: A test procedure for High Availability in Cloud Operating Systems, in: Vehicular Technology Conference, Las Vegas, USA, September 2–5, 2013, pp. 1–8.

47.

Brenner

Garbers

and Kapitza

, Adaptive and Scalable High Availability for Infrastructure Clouds, in: 4th International Conference on Distributed Applications and Interoperable Systems, Berlin, Germany, June 3–4, 2014, pp. 16–30.

48.

Malathy

Somasundaram

and Duraiswamy

, Performance improvement in cloud computing using resource clustering, Journal of Computer Science 9 (2013), 671–677.

49.

Chavan

and Kaveri

P.R.

, Clustered virtual machines for higher availability of resources with improved scalability in cloud computing, in: First International Conference on Networks and Soft Computing, Guntur, India, August 19–20, 2014, pp. 221–225.

50.

Saravanakumar

Geetha

Kumar

S.M.

Manikandan

Arun

and Srivatsan

, An Efficient Technique for Virtual Machine Clustering and Communications Using Task-Based Scheduling in Cloud Computing, Scientific Programming, 2021, 1–15.

51.

Cui

Zhang

Zheng

and Lou

, A cloud service resource classification strategy based on feature similarity, Journal of Networks 9 (2014), 2987–2993.

52.

and Kim

, A resource recommendation method based on dynamic cluste analysis of application characteristics, Cluster Computing 22 (2019), 175–184.

53.

Durgadevi

and Srinivasan

, Optimal resource discovery and dynamic resource allocation using modified hierarchal agglomerative clustering algorithm and bi-objective hybrid optimization algorithm, Asian Journal of Information Technology 15 (2016), 4464–4474.

54.

Meenakshi

Sirmathi

and Ruth

J.A.

, Cloud computing-based resource provisioning using k-means clustering and GWO prioritization, Soft Computing 23 (2019), 10781–10791.

55.

Shahapure

N.H.

and Jayarekha

, Replication: A technique for scalability in cloud computing, International Journal of Computer Applications 122 (2015), 8875–8887.

56.

Shahriar

Sharmin

Ahmed

Rahman

Boutaba

and Mathieu

, Diurnal Availability for Peer-to-Peer Systems, in: The 9𝑡ℎ Annual IEEE Consumer Communications and Networking Conference, Las Vegas, USA, January 14–17, 2011, pp. 619–623.

57.

Achache

Baaziz

and Sari

, The Impact of Data Mining and SaaS-Cloud Computing, in: 3𝑟𝑑 International Conference on Data Science and Applications, Istanbul, Turkey, June 25–28, 2020.

58.

Chen

Hua

Chang

Wang

Zhang

and Kong

, A Survey of Collaborative Filtering-Based Recommender Systems: From Traditional Methods to Hybrid Methods Based on Social Networks, in IEEE Access 6 (2018), 64301–64320.

59.

Shree

S.U.

and Basha

M.S.S.

, An exaustive survey of trust models in P2P network, International Journal on Web Service Computing 5 (2014), 1–12.

60.

Koutrouli

and Tsalgatidou

, Reputation-based Trust Systems for P2P Applications: Design Issues and Comparison Framework, in: Third International Conference onTrust and Privacy in Digital Business, Kraków, Poland, September 4–8, 2006, pp. 152–161.

61.

Visalakshi

N.K.

and Kuttiyannan

, Impact of normalization in distributed k-means clustering, International Journal of Soft Computing 4 (2009), 168–172.

62.

Jesi

G.P.

, PeerSim HOWTO: Build a new protocol for the PeerSim 1.0 simulator, 2005.

63.

Srati

Jinwala

D.C.

and Garg

, A survey of simulators for P2P overlay networks with a case study of the P2P tree overlay using an event-driven simulator, Engineering Science and Technology, an International Journal 20 (2017), 705–720.

64.

Rendón

Abundez

Arizmendi

and Quiroz

E.M.

, Internal Versus External Cluster Validation Indexes, International Journal of Computers and Communications, 2011.

65.

Hassani

and Seidl

, Using internal evaluation measures to validate the quality of diverse stream clustering algorithms, Vietnam Journal of Computer Science 4 (2017), 171–183.

66.

Wang

and Zhang

, On fuzzy cluster validity indices, Fuzzy Sets and Systems 158 (2007), 2095–2117.

67.

Kaya

, An algorithm for image clustering and compression, Turkish Journal of Electrical Engineering and Computer Sciences 13 (2005), 79–91.

68.

Davies

D.L.

and Bouldin

D.W.

, A cluster separation measure, IEEE Transactions on Pattern Analysis and Machine Intelligence 1 (1979), 224–227.

Hybrid fuzzy clustering to improve services availability in P2P-based SaaS-cloud

Abstract

Keywords

1. Introduction

Table 1 Summary of the main papers on improving availability in the cloud through replication, scheduling and fault tolerance mechanisms

3.1 System model

3.2 Methodology

3.2.1.1. Client profiles data structure

3.2.1.2. Feature normalization

3.2.2.1. Clients’ profiles data clustering

3.2.2.2. Virtual nodes and virtual layer formation

4.1 Experimental setting

Table 3 Configuration settings

4.2.1 Clustering evaluation criteria

4.3 Results and discussion

6. Conclusion

Footnotes

Author’s Bios

References

Table 1
Summary of the main papers on improving availability in the cloud through replication, scheduling and fault tolerance mechanisms

Table 3
Configuration settings