Phase-Aware Resource Tuning Using Behavioral Inference

Abstract

Technologies are disclosed for tuning resources of a processing unit based on behavioral phase-of-execution determinations. A phase-of-execution of code executing on a processing unit is inferred from processor-level execution events detectable at an architectural level. A tuning controller applies phase-specific adjustments to one or more hardware resources in real time with execution of the code. In some embodiments, tuning prepares for a predicted subsequent phase-of-execution and is applied prior to commencement of the predicted phase. Expected performance benefit may be evaluated by estimating the probability that a predicted phase is correct, the performance improvement if correct, and potential degradation if incorrect, with tuning applied only when an overall expected benefit is positive. Phase-specific adjustments may include modifying power gating, voltage or frequency levels, cache configuration, prefetching behavior, branch prediction parameters, or speculative-execution mechanisms. These techniques enable adaptive, phase-aware optimization of processing-unit resources without reliance on static knowledge of code structure.

Dependencies are inferred from the claim text.