Optimierung der Dienstauswahl in einem CORBA-Trader unter dem Aspekt dynamischer Lastverteilung

The growing performance needs and the possibility to use globally offered services via the internet made distributed systems become popular. A common approach to speed up the performance of a system is to distribute the offered service to several servers in order to balance the resulting load. A concept to find similar services offered by several servers is trading. A trader matches exported service offers and import requests by comparing the type and the properties of the considered services. This thesis evaluates an approach for an optimized service selection in a CORBA (Common Object Request Broker Architecture) based distributed environment. This is done by transparently integrating a load balancer component to an existing trader. Thus, the overall performance of the services offered by a distributed system can be increased. To make use of dynamic load balancing, monitoring and management components are necessary. For the sake of simplicity the developed management components are restricted to load monitoring. <p> A survey of existing concepts for trading or load balancing presented in this thesis shows that these are of theoretical nature or either restricted to load balancing or trading only. Just two projects which combine both aspects are known. In contrast to the architecture which is developed in this thesis, both projects lack of the possibility to compromise between the quality of the service offer and the desired small server load. <p> The presented trader provides an automatic load balancing of shared servers in a CORBA based distributed environment. To evaluate different load balancing strategies and load metrics a flexible load monitoring subsystem was developed. Server-side sensors are placed as CORBA compliant filters in the server stub generated from the IDL (Interface Definition Language) interface description. Intelligent local and global MIBs (Management Information Bases) are able to compute several simple and complex load metrics from the sensored data. During an import request the trader matches all exported service offers against the desired service type. The matching offers are scored based on the quality of their service offer property and their predicted load which is computed on the fly by the MIBs. A compromise between these weightable scores is calculated using a vector norm based rule which regards service property quality and load as components of a two dimensional quality vector. <p> This thesis focuses on the evaluation of different load balancing strategies and load metrics which are employable in the developed load balancing trader. The most important conclusion which can be drawn from the performance measurements is that simple dynamic load balancing strategies, like considering just the queue length, are best suited for systems with heterogeneous computing power and changing background load. This confirms the existing theoretical discussions stating more sophisticated dynamic strategies produce large overhead with no significant increase of load balancing quality. Moreover, in practice sophisticated strategies can be fooled by requests with non­predictable service times. Static load balancing strategies and strategies relying only on the traders knowledge are not suitable for open heterogeneous distributed systems found in real life because changing service times are not considered properly. The weights and vector norms used to compromise between trader and load balancer are dependent on the users' preferences as shown in several examples. The performance overhead in the trader due to the integration of dynamic load balancing is minimal if a caching strategy is used to transfer and update the monitoring information kept in the global MIB. <p> As the main contribution this thesis shows that integrating dynamic load balancing into a trader is very worthwhile, yielding a performance boost concerning the overall service time with little additional overhead to the trading time.