A Cross-Core Performance Model for Heterogeneous Many-Core Architectures

An accurate performance predictor to identify the most suitable core-architecture to execute each thread/workload in a heterogeneous many-core structure is proposed. The devised predictor is based on a linear regression model that considers several different parameters of the many-core processor architectures, including the cache size, issuewidth, re-order buffer size, load/store queues size, etc.. The devised predictor is easily integrated in most system schedulers, providing the ability to periodically determine whether a certain thread is running in the most efficient core-architecture. The obtained experimental results show that the devised model is able to identify the correct core-architecture in a large majority of the cases, leading to average performance differences as low as 7% when compared with an oracle scheduling solution.