New content-aware request distribution policies in web clusters providing multiple services

Abstract

Due to the explosive growth of the Internet and increasing service demands from all around the world, the cluster-based system that consists of one request-dispatching server and several requesthandling servers has become a cost-effective way to serve the huge amount of service demands. Nowadays, Web servers have to handle more complex types of requests since requests from clients may be mixed with dynamic Web pages, database processing, or multimedia stream data. Therefore, a Web cluster should be designed with intelligent request dispatching policies for supporting various types of service requests.In this paper, we have proposed two new content-aware request distribution policies named Locality-Aware Request Distribution with Replication and Classification (LARD/RC) and Grouped Client-Aware Policy (GCAP) to dispatch requests efficiently in Web clusters providing multiple types of services and running in homogeneous or heterogeneous environments. We have implemented our proposed policies in the LVS-CAD web cluster that can efficiently perform content-aware request dispatching. Performance evaluation shows that our proposed LARD/RC and GCAP policies could dispatch requests of different types to proper back-end servers in a more efficient way to utilize system resources than the other existing policies in both homogeneous and heterogeneous environments.

A New MapReduce Framework Based on Virtual IP Mechanism and Load Balancing Strategy

“…Load balancing strategy [8] mainly includes RoundRobin Scheduling (RR), Weighted Round-Robin Scheduling (WRR) [9], Least-Connection Scheduling(LC) [10], Weighted Least-Connection Scheduling(WLC) [11], Locality-Based Least-Connection Scheduling(LBLC) [12], Locality-Based Least-Connection with Replication Schedul-ing(LBLCR), Destination Hashing Scheduling(DH), Source Hashing Scheduling and so on. Aiming at practical situation or experimental questions, algorithms can be restructured to optimize job assignment, such as appending dynamic detection based on the state of node, or adding algorithms based on some effective math conclusion into load balancing strategy.…”

An up-to-date survey in web load balancing

“…There are other proposals that also include a set of the most accessed files to be cached in all the servers [19,20,32,51,62,74]. A prefetching policy named Time and Access Probability-based Prefetch (TAP 2 ) is proposed by Park et al in [62].…”

“…Results show that CWARD/FR outperforms CWARD/CR. Chiang et al, in a later paper [20], propose two other content-aware policies based on CAP (described in the next subsection) and LARD, named Grouped Client-Aware Policy (GCAP) and Locality-Aware Request Distribution with Replication and Classif ication (LARD/RC), respectively. In these new policies, the authors improve the treatment of dynamic Web pages, database processing and multimedia stream data, that might easily overload the backend servers, especially in heterogeneous environments.…”

Autonomous learning of load and traffic patterns to improve cluster utilization

“…Locality-based least connection with replication (LBLCR) scheduling that comes with Linux is designed to adapt to different traffic patterns in a cluster [19]. Linux Virtual Server (LVS)-based load balancer [9,26] distributes client requests to back-end servers based on the number of connections between the LVS load balancer and the back-end servers. The decision as to which server to send the request is made based solely on the number of connections, or threshold, regardless of the implications of each connection.…”