Analysis and evaluation of performance issues of parallel software on multi-core processors

For decades, the processor manufacturers have attempted to achieve performance gains by increasing the clock frequency on single-core processors. But physical problems-such as the high power dissipation-lead to the release of the first multi-core processors on the market in 2005. To benefit from the multi-core architecture, parallel programming is required. However, this programming model requires a different approach and is associated with certain risks and pitfalls. This paper focuses on modelling of certain test scenarios for two common multi-core specific problems, namely oversubscription and false sharing. Various simulations and tests offered solutions and design patterns to avoid such problems. Results have shown that the problems have a fatal impact on the execution time, so that the performance gain on the multi-core system is nearly nonexistent. Thence, any software developer must have in-depth knowledge of the used hardware and software to benefit as much as possible from multi-core architectures.