Untersuchungen zur Kommunikation von Verteilten Betriebssystemen über Hochgeschwindigkeitsnetzwerke

摘要

In the course of improving the performance of computing systems, several different architectures have been developed, among which during the last years the cluster architecture proved to be the dominating one. This architecture combines the advantage of high computational power with scalability and reliability, but it comes at the cost of firstly more complex management of the distributed components and secondly slower communication due to the interconnection network. In order to simplify management and handling of the distributed computing resources, operating systems have been developed which hide the distributed nature of the components from the user, providing him the image of a singular centralised system. The problem of communication bottlenecks in distributed architectures is usually tackled by special high-speed networks, which can be accessed by applications using standard programming interfaces, but could not be used for clusterwide inter-process communications by freely available Linux-based distributed operating systems. Within this thesis, two implementations of a software layer that allows free Linux-based distributed operating systems - especially their representative Kerriged - to use high-speed networks for communication have been developed. One of the main aims during the development was to support a large number of high-speed networks and a wide spectrum of applications. Furthermore, two additional concepts were developed, which allow for more direct communication and therefore are expected to deliver even higher performance, sacrificing the high flexibility concerning applications and restricting the usage to applications that use the TIPC network protocol for communication. Measurements demonstrate that by using the developed software layers, communication performance of the distributed operating system Kerrighed can be greatly improved. Throughput shows up to three times higher levels compared with usage of Gigabit Ethernet, while latency can be decreased to half that of Gigabit Ethernet.