Research On Key Technologies Of Middleware Based Load Balancing For Heterogeneous Distributed Environment

With the development of the network, the distributed applications keep increasing day by day. In order to service the increasing online clients those transmit a large, often bursty, number of requests and provide dependable services with high quality constantly, lots of systems begin to make use of the redundant servers connected by the high speed network to form the server clustering to serve the peak time of the client requests efficiently. Furthermore, to the key business and military applications, the systems should provide transparent and adaptive support for the regular running. Therefore, under this background technologies for load balancing have emerged as the times require and have become one of the hot points in the research of distributed computing.Currently, load balancing mechanisms can be provided in a distributed system in network layer, operation layer, application layer and middleware layer. Usually, application based load balancing is not transparent to the clients for the applications should be revised to complete the load balancing by themselves so that it is more complex to develop and maintain the programs.Furthermore, network based load balancing and operation system based load balancing don't support the load metric defined by the applications, so users may pay no attention to the selection and customize of the load balancing strategies which may be better. Therefore, this thesis focus on the middleware based load balancing which can use the advantages of middleware such as transparency, extensibility and reflective. It can not only make the load balancing to the clients and servers as transparent as possible but also support algorithms customized by the users to adapt to the need of different applications and improve the scalability and adaptability of the systems.Based on the fruit and deficiency of the existing works, aiming to support high performance, better availability and scalability for the more and more complex distributed applications, this dissertation conducts an in-depth study which focuses on middleware based load metrics, load monitoring, load balancing algorithms, overload control and autonomic replica management in the distributed heterogeneous environment. Then an adaptable, transparent and scalable load balancing service is brought forward. The main contents of this dissertation are as follows:1. Scalable load metrics and load monitoring mechanisms. Based on the compare and analysis of existing load metrics, this paper brings forward a component based load metric definition approach by which users can define private load metric according to the need of special applications. At the same time, to eliminate the redundant overhead introduced by the traditional per object load monitoring when multiple kinds of replicas exist in the same host, this paper put forward a multiple granularity load monitoring approach MGLMA. This approach makes use of the agent to organize the replicas having the similar load metric into the virtual group and actual load monitoring is performed by the agent. Therefore, the redundant load monitoring can be eliminated and granularity of load monitoring can be adjusted flexibility. The experimental results show that, this approach can diminish the extra overhead and improve the scalability of the system when multiple kinds of replicas are existing in the same host.2. Load balancing algorithms. On the one hand, based on the classified distribution of existing load balancing algorithms this paper puts forward a component based architecture which can help the users to expand as well as reconfigure the algorithms dynamically. On the other hand, this paper put forward a triggered round robin algorithm TRR based on the round robin algorithm and the weighted round robin algorithm for the distributed heterogeneous environment. This algorithm has the same advantage with the weighted round robin algorithm for it can adapt to the back end resources dynamically. However it doesn't report load values to the load balancer periodically and the communication is performed just when needed. The experimental results show that, this algorithm has lower communication overhead when compared to the weighted round robin algorithm and less little possibility of overload when compared to the general round robin algorithm. Besides these, this paper strengthens the least load algorithm and bright out the location based least load algorithm LLL. The experimental results show that, this algorithm has less communication and computation overhead. Based on the above research this paper analyzes the performance of different algorithms. The experimental results show that, the adaptive algorithms and the non-adaptive algorithms have the similar throughput and response time when using similar workloads while the adaptive algorithms may have better performance when using heterogeneous workloads.3. Overload control and load rebalancing. In existing load balancing service the workloads are controlled in a damping way and the damping factor is configured statically and it can't be adjusted dynamically. Therefore this paper puts forward a machine learning based overload control approach MLOCA to control the damping factor in a more flexible and dynamic way to improve the efficiency of overload control. The experimental results show that, the approach can flatten the peak of the workloads and improve the efficiency of overload control. Secondly, aiming at the problems existing in traditional replica management mechanisms such as low resource utilization and failure of hot pot services this paper bring up an autonomic replica management approach ARMA to make the system can control the creation and the elimination of the replicas autonomically and realize balanced resource allocation among different kinds of replicas by control the number of the replicas. The experimental results show that, this approach can improve the throughout, efficiency and stability of the systems whenever the services have priorities or not. 4. Design and implementation of the system. Based on the studies of the key technologies stated above and the object oriented distributed computing platform StarBus, this paper goes on with the implementation issues of scalable middleware based load balancing service StarLB to support load balancing transparently and efficiently for the applications in the distributed heterogeneous environment.