Predicting Thread Profiles across Core Types via Machine Learning on Heterogeneous Multiprocessors

Given that energy consumption has become one of the most important issues in computer systems, Heterogeneous Multiprocessors (HMPs) have been introduced, where large high performing and small power-efficient cores can co-exist on the same platform and share the processing of the workload. Clearly, the concept is the same whether it is multiple processors on a board or a chip multiprocessor with several cores on a chip. With the advent of HMPs, thread scheduling becomes much more challenging, while having to deal with thread to processor-type mapping. In particular, it is important that the operating system is able to understand the workload behavior when a thread is to be migrated to a core of a different type. In this paper, we describe a thread characterization method that explores machine learning techniques to automate and improve the accuracy of predicting thread execution across different processor types. We use hardware performance counters and use machine learning to predict performance when moving a thread to another core type on heterogeneous processors. We show that our characterization scheme achieves higher structural similarity (SSIM) values when predicting performance indicators, such as instructions per cycle and last-level cache misses, commonly used to determine the mapping of threads to processor types at runtime. We also show that support vector regression achieves higher SSIM values when compared to linear regression, and has very low (1%) overhead.