Figure 1 - uploaded by Krzysztof Zatwarnicki
Content may be subject to copyright.
Source publication
This paper presents a novel algorithm for distribution of user requests sent to a Web-server cluster driven by a Web switch. Our algorithm called FARD (fuzzy adaptive request distribution) is a client-and-server-aware, dynamic and adaptive dispatching policy. It assigns each incoming request to the server with the least expected response time, esti...
Contexts in source publication
Context 1
... various academic and commercial parallel or distributed applications [2]. As the Internet started to be bigger and more complex, the load balancing performance strategy has also been widely deployed for Web systems to address the needs of highly trafficked and mission-critical Web sites. In this paper, we are dealing with Web-server clusters ( Fig. 1) consisting of multiple server nodes, distributed within a local area and equipped with a mechanism to spread client requests among the nodes. Web servers in a cluster work collectively as a single Web resource in the network. Typical Web cluster includes a Web switch that has ability to distribute user requests among Web servers in a ...
Context 2
... illustrated in Fig. 1, the Web switch is placed in front of the set of Web servers using a two-way architecture. All URL requests arrive through the Web switch and the resulting resource is sent back also through the Web switch. Although a client issues one request at a time for a Web page, their action usually causes multiple client-to-server interactions ...
Context 3
... target server is chosen in the following way. When x i request arrives, each Model of Server calculates its own estimator of the response time t i *s , s=1,..,S assuming that this request is to be serviced by the given server, and making use of current server load information (c,a,d) s , s=1,..,S and the knowledge about the request x i (type and size of the object). ...
Context 4
... the performance for FARD is much better with respect to other policies. It is worth noticing that when a Web server handles a big number of clients, the response time may increase to be rather significant, so the clients become impatient, and therefore we should try to make individual response times as short as possible. The results given in Fig. 10 confirm that FARD policy operates in that way. Fig. 11 shows "light requests" to perform under FARD policy better as compared with LARD algorithm and far better as compared with WRR algorithm, while large objects, those greater than 50 KB, perform only negligibly worse under FARD, compared to other policies. The average object size in ...
Context 5
... to other policies. It is worth noticing that when a Web server handles a big number of clients, the response time may increase to be rather significant, so the clients become impatient, and therefore we should try to make individual response times as short as possible. The results given in Fig. 10 confirm that FARD policy operates in that way. Fig. 11 shows "light requests" to perform under FARD policy better as compared with LARD algorithm and far better as compared with WRR algorithm, while large objects, those greater than 50 KB, perform only negligibly worse under FARD, compared to other policies. The average object size in Profile 6 was 39 KB, which some authors assume to be ...
Context 6
... requests" to perform under FARD policy better as compared with LARD algorithm and far better as compared with WRR algorithm, while large objects, those greater than 50 KB, perform only negligibly worse under FARD, compared to other policies. The average object size in Profile 6 was 39 KB, which some authors assume to be typical for the Web. Fig. 12 presents the result for heterogeneous cluster. It is clear that for such configurations of Web-server cluster, FARD significantly outperforms other policies. We may also conclude that FARD outperforms other policies particularly for smaller objects. The maximum throughput is limited for FARD to 5,000 connections per second, whereas ...
Similar publications
![Table 2 . 90-percentile of response time [s]](profile/Riccardo-Lancellotti/publication/237515591/figure/tbl2/AS:644268670599176@1530617029488/90-percentile-of-response-time-s_Q320.jpg)
The advent of the mobile Web and the increasing demand for personalized contents arise the need for computationally expensive services, such as dynamic generation and on-the- fly adaptation of contents. Providing these services exacer- bates the performance issues that have to be addressed by the underlying Web architecture. When performance issues...
Citations
... In existing work [2], L.Borzemski and K.Zatwarnicki has presented a novel algorithm for load distribution of user requests sent to the clustered web server systems driven by a web switch. They proposed an algorithm called Fuzzy Adaptive Request Distribution is a client and server aware, dynamic and adaptive dispatching policy. ...
... Another group of intelligent adaptive approaches using fuzzy neural models was proposed by Zatwarnicki. These are solutions enabling global distribution between server rooms located in different geographic locations, e.g., GARD [38] and GARDIB [7], and also local approaches such as FARD [39] and FNRD [6] strategies. All systems presented in the solutions work in a way that minimizes the response time for each individual HTTP request. ...
Cloud computing systems revolutionized the Internet, and web systems in particular. Quality of service is the basis of resource configuration management in the cloud. Load balancing mechanisms are implemented in order to reduce costs and increase the quality of service. The usage of those methods with adaptive intelligent algorithms can deliver the highest quality of service. In this article, the method of load distribution using neural networks to estimate service times is presented. The discussed and conducted research and experiments include many approaches, among others, application of a single artificial neuron, different structures of the neural networks, and different inputs for the networks. The results of the experiments let us choose a solution that enables effective load distribution in the cloud. The best solution is also compared with other intelligent approaches and distribution methods often used in production systems.
... In the previous work [5][6][7][8] a research on intelligent fuzzy-neural distribution methods minimizing service time in cluster-based and cloud-based Web systems has been presented. In the field of cloud systems, an effective HTTP request distribution system using two-layer architecture, in which decisions were made on two independent levels by web switches learning mutual behavior has been proposed [9,10]. ...
... Those are the solutions enabling global distributions among server rooms located in different geographical locations e.g. GARD [6] and GARDIB [8] and local approaches like FARD [5] and FNRD strategies [7]. All the systems work in this way to minimize the response time for each HTTP request separately. ...
... The new strategy is called Two-Level Fuzzy-Neural Request Distribution, or simply TLFNRD. The main difference between methods and algorithms presented in the works [5][6][7][8][9][10] and the TLFNRD strategy is that the previously proposed Web switches and brokers make the distribution decision only on one level (Fig. 2a). Those switches and brokers are estimating service times using the fuzzy-neural mechanism for all executors servicing HTTP requests (e.g. ...
Abstract Cloud computing Web systems are today the most important part of the Web. Many companies transfer their services to the cloud in order to avoid infrastructure aging and thus preventing less efficient computing. Distribution of the load is a crucial problem in cloud computing systems. Due to the specifics of network traffic, providing an acceptable time of access to the Web content is not trivial. The utilization of the load distribution with adaptive intelligent distribution strategies can deliver the highest quality of service, short service time and reduce the costs. In the article, a new, two-level, intelligent HTTP request distribution strategy is presented. In the process of designing the architecture of the proposed solution, the results of earlier studies and experiments were taken into account. The proposed decision system contains fuzzy-neural models yielding minimal service times in the Web cloud. The article contains a description of the new solution and the test-bed. In the end, the results of the experiments are discussed and conclusions and presented.
... Another group of intelligent adaptive approaches, using fuzzy-neural models, was presented in the articles of the author and the research group. There have been proposed strategies minimizing response times of HTTP requests in cluster-based web systems [4,20] and also solutions enabling global distribution among regions [5,19]. Motivation Our latest work was devoted to the problem of adapting the solutions designed for cluster-based web systems to work effectively in the web cloud. ...
... All of the information is available within the TLFNRD broker. 4 Estimation of HTTP request response time , . . . , , . . . ...
Web Cloud systems are very popular today. One of the main problems in cloud computing today is the better use of distributed resources and applying them to achieve higher throughput. To solve those problems load distribution mechanisms are implemented. A two-level decision HTTP request distribution strategy working in a one-layer architecture is presented in the article. The proposed solution uses fuzzy-neural models yielding minimal service times in the web cloud. The article contains the results of experiments in which a new solution is compared with intelligent request strategies used in two-layer architectures. Discussion of results and final conclusions are presented at the end.
... Adaptive learning methods keep learning and updating system parameters in real-time while the system is working. The request distribution strategy of Fuzzy Adaptive Request Distribution (FARD) [24][25][26] used traditional error back propagation algorithm in training the learning system, so as to predict the response time of each back-end node in real-time. Reference [27] used an adaptive learning BPNN to get hold of the relationship between server nodes' percentage of CPU usage, processor clock speed, memory utilization and capacity, network bandwidth utilization, server's NIC bandwidth and their response time. ...
... A separate group of distribution algorithms are adaptive algorithms which can adapt to a changing environment. The most well-known are AdaptLoad [13] which distribute the request on the base of the size of requested objects, Adaptive Load Balancing Mechanism (ALBM) [8], which take into account the load of the server applications and finally two algorithms, proposed by the author, minimizing request response time, Fuzzy Adaptive Request Distribution (FARD) [3] and Local Fuzzy Neural Requests Distribution (LFNRD). A good survey of different distribution algorithms can be found in [9]. ...
This work presents the application of two adaptive decision making algorithms in Web switch controlling operations of cluster-based Web systems. The proposed approach applies machine learning techniques; namely a fuzzy logic and neural networks, to deploy the adaptive and intelligent dispatching within locally distributed fully replicated Web servers. We present the adaptive control framework and the design of Web switch applying our conception. We demonstrate through the simulations experiments the difference between our algorithms and compare them with the most popular and reference distribution algorithms.
... TCPsplicing FARD [8] 03 ...
... There are other proposals that do include monitored information obtained from the Web servers of the cluster to take distribution decisions. This is the case of Fuzzy Adaptive Request Distribution (FARD) [8]. This work of Borzemski and Zatwarnicki describes a content-aware load balancing mechanism that estimates the response time of each request in every server of the cluster using a fuzzy estimation mechanism. ...
This survey is an up-to-date state-of-the-art in Web load balancing mechanisms that includes all the possible classifications
and focuses on the advantages of using load balancing solutions to increase the performance of the Web system. A general description
of the Web load balancing solutions is included and organised by differentiating the OSI protocol stack layer the load balancing
is based on. The most important request distributing polices that are proposed in the literature are also included. This article
summarises all previous surveys on the Web load balancing subject and updates the state-of-the-art with the most recent load
balancing proposals.
... Leszek Borzemski [27] proposed a novel fuzzy adaptive request distribution (FARD) algorithm for dispatching the user requests and selecting target server in a web cluster system. Dispatching decisions are dynamic and are based on current knowledge about the server state. ...
... The solutions for web clustering were given by methods designed using neural and fuzzy systems. In case of the method discussed in [27], it also considered neural network as part of its selection process, in addition to the fuzzy system designed for clustering. Neural network approach was designed to cluster users according to their request patterns to improve the access performance. ...
Web caching has been used extensively to enhance content delivery to the clients by minimizing -client- -observed latency, reducing network bandwidth usage and improving scalability of the network. Caching performance can be improved by designing good replacement policies, prefetching techniques, clustering of web users and proper placement of proxy caches in the network. In this paper, we discuss the various approaches that were designed based on neural networks, genetic algorithms and fuzzy logic to optimize the performance of web caching. The approaches discussed here proved to be more effective in solving the problems as compared to the conventional techniques that were used earlier in this problem domain. Neural networks and evolutionary algorithms can be considered for further exploration in the various issues related to web caching and content delivery.
... In our previous work we have developed an adaptive content-aware request distribution algorithm called FNRD for dispatching HTTP requests in a local Web cluster (Borzemski and Zatwarnicki 2003). We showed by means of extensive tests and evaluations (Borzemski andZatwarnicki 2005, 2006) that FNRD algorithm is useful and outperforms other competitive dispatching policies, including the state-of-the-art content-aware ''reference'' algorithms such as CAP (Content Aware Policy; Casalicchio and Colajanni 2001) and LARD (Locality Aware Request Distribution; Pai et al. 1998), as well as very popular content-blind RR (Round Robin) policy (Cardellini et al. 2002). ...
... An adaptive decision making framework for Web request distribution was first proposed by Borzemski and Zatwarnicki (2003). This concept was invented to support HTTP dispatching within a Web server cluster. ...
... Any local request distribution policy can be used at local level service. In this article we use FNRD algorithm (Borzemski and Zatwarnicki 2003, 2005. ...
Today's knowledge and content Web-based technologies call for adaptive and intelligent algorithms for accessing the resources, especially to respond to Semantic Web, as well to performance challenges. This work presents the application of two machine learning techniques in an adaptive decision making framework, namely a fuzzy logic and neural networks, to deploy the adaptive and intelligent dispatching algorithm for resource requesting within a geographically distributed fully replicated Web sites. Our approach and algorithm can be used in content distribution and delivery networks where due to content replication each Web server can respond to the client request. Intelligent redirection mechanism redirects the request from the originating site's server to the “best” content source to deliver the content in the fastest manner. First, we introduce a broker-based system architecture. Next, we develop a new global adaptive intelligent request distribution algorithm. Our algorithm called GARDiB by a means of the fuzzy-neural techniques redirects each incoming request to local service with the least expected response time, which takes into account, the transfer time of the request and the response, as well as the time elapsed on the local service that responds to the request. Our simulations showed the effectiveness of the proposed adaptive distribution.
... in addition of cost, a single server machine can handle limited amount of request and cannot scale up with the demand. This will be clearer when we know 40% latency, as perceived by users, is causing on a web site side ( Cardellini & al., 2002) . The need to optimize performance of web sites is producing a variety of novel architectures. ...
... Various kind of systems such as Magic router(Anderson & al., 1996) , Distributed packet rewriting (Aversa & al., 2000), Cisco local director (Cisco, 2004) , LVS (lvs, 2004), IBM network dispatcher(Hunt & al., 1998) , ONE-IP (Damani & al., 1997) and Dynamic feedback (Kerr., 2004) , uses layer-4 web switch. One of frequently used algorithms in layer-4 web switch is Weighted Round Robin (WRR), that associates to each server a dynamically evaluated weight that is proportional to the server load state (Cardellini & al., 2002). Layer-7 web switch can establish a complete TCP connection with the client and use OSI session layer information, such as session identifiers, file size, file type and cookies (Aron & al., 1999) , in addition to layer 2,3 and 4 information, prior to decision making. ...
In this paper we propose a new scheduling policy namely IRD (Intelligence Request Dispatcher) for web switches operating at layer-7 of OSI protocol stack in a web server farm .we classify dynamic and static requests. A hybrid Neuro-Fuzzy and LARD like approach is used to make route decision of dynamic and static requests separately. IRD select server with lower response time in an adaptive dispatching policy. In particular, we used the ANFIS (Adaptive Neuro Fuzzy Inference System) methodology to build a Sugeno fuzzy model to assigns each incoming dynamic request to the server with the least expected response time. This estimation is based on expected impact of each client's individual request on server's resources. this resources are , number of connection, CPU and disk usage .This parameters used to evaluate a load-weight of each server in a fuzzy manner and retrieved through feed-back. For Static requests an algorithm used to improve the cache hit rate in Web cluster nodes according to their load-weights. It's goal is to improve load sharing in web clusters with dynamic and static information. A prototype implementation results confirm that proposed algorithm is more effective than representative request distribution algorithms, specially for dynamic requests that requires sophisticated policy and more service time than static information.
