The increasing prevalence of multi-core processors means that applications must be multi-threaded to effectively use available hardware resources. The number of cores and threads to be employed in order to optimize the performance for a specific architecture is not always clear--using too few might constrain the inherent parallelism, while too many might cause high synchronization overhead.

This paper addresses the problem of multi-core performance prediction. First, the authors benchmark the performance of various programs of the Java Grande suite, on two Sun multi-processor architectures using Solaris containers to control the number of central processing units (CPUs). Then, they hypothesize the logarithm of the performance to be a linear function of the logarithm of the number of CPUs and the logarithm of the number of threads, and apply linear regression analysis to determine the optimal prediction function for each program and architecture. Next, they develop a neural network model that takes the number of CPUs and the number of threads as inputs, and predicts performance. While the linear regression model often generated acceptable predictions, the neural network model gave, in all cases, after suitable training, very close approximations of the actual performance.

While the neural network model does not provide much human insight, it may find its application as a black box in the auto-tuning of parallel applications. An interesting side effect of the analysis was the observation that optimal performance is achieved with a one-to-one mapping of threads to cores; that is, the performance ultimately depends on one parameter only.