2026 , Volume 31, № 1, p.106-120

The aim of the work is to optimize the number of phases of the processor pipeline according to the criterion of minimum application execution time and analyze the conditions for the feasibility of pipelining.

The study of the pipeline computing process is based on a model of the central processor considered as a homogeneous pipeline processing a heterogeneous flow of commands with heterogeneous probabilities of rebooting the pipeline. Rebooting occurs when branch prediction failures in various sections of the software algorithm.

The dependence of the execution time of an application of a given size on the known complexity of individual commands and the overhead of saving and extracting intermediate results from processor control structures at individual pipeline phases is found. Based on the obtained dependence, the ratio for the optimal length of the processor pipeline is calculated analytically, which is found to be consistent with known special cases. Conditions for accelerating calculations in the form of inequalities are formulated for the parameters of the processor pipeline and the characteristics of applications.

The obtained ratio for the optimal length of the processor pipeline with dynamic command execution allows assuming that the reboot probability is nonzero only in the sections of the application algorithm which contains conditional operators and commands that follow them within the sliding window (buffer) of processor commands. It allows more accurate accounting for the reasons of the processor pipeline reboots and optimizing the execution time of applications.