“Load Management Systems for Component-based Middleware”

This thesis proposes a new approach to the design of reflective load management services for middleware, tackling the main problems in existing load management services. The system is designed using a modular architecture. The two key benefits of the approach are that modules can be dynamically activated and deactivated as required, enabling the minimisation of the overhead introduced by the system, and that new modules with enhanced functionality can easily (and dynamically) be introduced into the system. The system comprises of a load monitoring module, a load prediction module, a load evaluation module and a load distribution module. Each module has a clearly defined role in the system and a well-defined interface. The load evaluation module offers the possibility of dynamically changing the distribution algorithm. The modularity of the system is further extended to the monitoring, workload prediction and load distribution components, so that new monitors and algorithms can be added at runtime. A novelty of the proposed approach consists in the inclusion of QoS in the load management system, thus making it transparent to the managed applications. This approach offers increased flexibility and reusability because QoS can be added to existing (non QoS-aware) applications without the hurdle of changing the code. The response time metric is used for QoS level differentiation. An important characteristic of our load management service is that it is transparent to distributed application developers. The design of the load management system ensures high availability by including a simple load distribution mechanism in the distribution module. Another novelty of this approach is the automatic selection of the optimal load distribution algorithm at runtime, according to current system state and workload. It is considered that the most important performance metric for system performance is the response time. An important achievement of such a reflective load management service is that it adapts itself at runtime to workload/environment changes without user intervention. A simulation model was created to evaluate existing load distribution algorithms. The model was extended to offer simulation scalability (e.g. the number of servers can be easily changed) and to support the evaluation of the newly proposed load distribution algorithm. The influence of the workload on the performance of the distribution algorithm was also investigated. The research approach employed carrying out an extensive literature survey in order to identify the main problems in existing load management services. These problems represented the ground for the framework proposed in this thesis. It is beyond the scope of this thesis to validate the entire framework thus only the key elements of the framework have been investigated in detail and validated using simulations.