This paper presents an online controller for regulating the throughput of instruction-sequences in multicore processors using dynamic voltage-frequency scaling. The proposed control law comprises an integral controller whose gain is adjusted online based on the derivative of the frequency-throughput relationship. This relationship is modeled as a stochastic DEDS having no analytic functional form, and hence its derivative is estimated by Infinitesimal Perturbation Analysis (IPA). However, the DEDS is multi-class and hence the IPA derivative is biased.

Biasedness of IPA is a common problem in multi-class DEDS which has hindered the development of IPA as a general tool for practical applications. However, recently it has been suggested that as long as the relative bias has certain upper bounds, optimization algorithms and control laws can still converge to optimal or near-optimal parameters. The purpose of this paper is to demonstrate this point for the aforementioned problem of throughput regulation, thereby suggesting the potential emergence of a new class of effective control laws in computer architectures.

Throughput Regulation in Multicore Processors via IPA [PDF]