Benchmarking Framework for Performance in Load Balancing Single System Image

Single System Image, as a distributed operating system for nodes in computer clusters has become a widely adopted clustering solution due to its complete transparency of the resource management and ease of use. An important design consideration for this environment is the load allocation and balancing which is usually handled by an automatic process migration daemon. Thus, the implementation of such mechanism becomes an important design consideration in the distributed operating system. There is an essential need for a benchmark framework for the Single System Image clusters due to the wide range of implementation and the need for identifying the performance and behaviour of the system. The benchmark framework will enable the researchers to investigate both relative behaviour and performance of the Single System Image clusters, as well as provides the ability to compare such systems. In this work, a carefully designed benchmark framework had been proposed to study and evaluate the performance of the load balancing single system image. The performance metrics, which takes into account the speed, nodes, network, and behaviour of the system, were formulated. The benchmark framework allows the determination of the performance degradation factors associated with system implementation and configuration. This framework has been utilized to assess the performance characteristics of an existing and successful Open Source load balancing SSI system, OpenMosix. The benchmark framework provides an understanding of how the SSI system responds under varying conditions and manages to characterize the limitation of the information dissemination algorithm of OpenMosix. The information dissemination daemon had also been improved. The performance of the improved strategy had been validated by comparing it with the original system. Finally, the results from the tests were combined into a single figure of the performance behaviour. The experimental results obtained from the benchmark framework showed that the numbers of nodes affect the performance of the SSI cluster; this could be regarded as an important factor of performance decaying. The number of nodes can affect the performance by adding extra costs including, but not limited to, network traffic, load balancing time, and overhead. The performance of any SSI cluster can be enhanced by improving any or all of the above factors. The improved load balancing strategy shows a visible performance gain with more than five nodes. At eight nodes, a gain of nearly 50 seconds runtime, 16.13 speedup, and 12.27 % efficiency have been successfully achieved