Machine learning performance prediction model for heterogeneous systems

Recent advances in machine learning techniques and specialized hardware has enabled a resurgence in the interest and applicability of powerful artificial neural network based prediction systems. However, as of yet no significant leaps have been taken towards applying machine learning in heterogeneous scheduling in order to maximize system throughput. As heterogeneous systems become more ubiquitous, computer architects will need to develop new CPU scheduling approaches applying novel techniques capable of exploiting the diversity of computational resources. However, non-heterogeneous aware schedulers like the current Linux Completely Fair Scheduler (CFS) [3] cannot take advantage of diverse system resources. Heterogeneous scheduling approaches have been previously proposed by V. Craeynest et al. [6] and Markovic et al. [2] which intend to provide full and fair utilization of the different hardware resources for all threads. These approaches yielded significant performance benefits compared to the CFS on heterogeneous architectures. In this extended abstract, we describe a novel performance prediction model which is the first of its kind to utilize machine learning performance predictors at the granularity of scheduling quanta. We then highlight how a heterogeneous system scheduler may be improved by the addition of this model. The prediction model is composed of the following: