Modelling the Performance of Processors in Heterogeneous Computing Environments
Robert Higgins · 2011
Accurate performance models of processors are essential for efficient heterogeneous parallel or distributed computing. Characterising the performance of a processor at a particular operation is a challenge. Simple models that are easy to construct and use do not represent enough of the detail of a processor’s performance to provide high efficiency in all conditions. However, detailed models are challenging to construct and more difficult to use in a data partitioning algorithm or task scheduler. In this thesis the construction and use of a detailed performance model is investigated. This model is titled: the Band Performance Model (BPM). It is an evolution of the Functional Performance Model (which expresses the speed of a processor as a function of the size of a specific task that it operates on). The BPM describes the processor speed not as a single valued function but as a function with a range of possible speeds based on a prediction of CPU availability. In this way the model encapsulates inherent variability in the performance of the processor. Variability which is caused by the concurrent execution of other processes. The Band Performance Model is formulated and a number of methods are described that use the model for partitioning of problems. The efficiency of the partitions is demonstrated. The cost of construction of such a model is addressed by a novel optimised building procedure. This procedure is implemented in a software tool which is demonstrated in detail. The tool is used to build performance models for a non-synthetic distributed application and these models are integrated with a Grid middle-ware, where significant performance increases are shown as a result of their use.