Active Resource Management in Middleware and Service-oriented Architectures

Abstract: In this thesis, we study load balancing and overload control in two distributed architectures: Distributed Object Computing (DOC) systems and Services-oriented Architecture (SOA) systems. While DOC and SOA have differences in terms of implementation, they are similar from a performance viewpoint. In both DOC and SOA, users access software applications distributed on multiple processing nodes. The applications are composed of intercommunicating objects or services. During execution, one or more of the software resources, distributed on multiple processing nodes, are needed. The complex interaction patterns that arise make overload control and load balancing hard. The first part of this thesis concerns load balancing in DOC systems. The first two papers evaluate request based load balancing in TINA and IN/CORBA respectively, comparing algorithms in terms of efficiency and overhead. The second paper also touches upon using lightweight agents to distribute load information in the network. The following two papers investigate two approaches to allocating objects to nodes. The allocation problem is a hard problem that is solvable in reasonable time only for small systems, however, its solution is crucial for efficient resource utilization. The first paper considers object migration algorithms as a way to dynamically distribute load. Two migration algorithms are are presented and evaluated. In the second paper, the static allocation problem is approached by applying the Cross Entropy Method, a stochastic optimization method, to the problem. The method is shown to be able to find good solutions without knowledge of system structure. The second part of this thesis concerns overload control in DOC and SOA. First comes an investigation of overload control in a real-time CORBA system. Second, request and session based admission control schemes in e-commerce web servers are compared. The session based mechanism is shown to be better in terms of user-experienced performance. The evaluation is continued in an appendix. Third, service protection during transient overload conditions in web services systems is studied. A fine-grained admission control framework is proposed and is shown, using deterministic flow models, to efficiently protect services that are not subject to overload.