Machine Learning for timing estimation

Estimating program execution time is a key but challenging task, further complicated by the growing complexity and insufficient documentation of modern processor architectures. While traditional methods like cycle-accurate simulators are precise, they are time-consuming and demand an in-depth understanding of the processor's architecture. A new data-driven approach utilizing machine learning techniques has been developed to address these limitations. However, while existing machine learning models offer rapid estimations, they are primarily tailored for simpler architectures with constant instruction timings. This thesis aims to develop new machine-learning methods for complex, undocumented processors by introducing context awareness into timing models based on machine learning. A novel approach treats instruction sequences like natural language and employs advanced machine learning algorithms such as Long Short-Term Memory networks and Transformers. This allows the model to consider complex features such as cache and pipeline effects, improving the accuracy for both worst-case and average-case execution times.

L’accès du public (hors IRISA) à cette soutenance est contraint à une inscription préalable obligatoire auprès de gaelle [*] tworkowskiinria [*] fr – L’accès ne sera pas autorisé sans inscription préalable. Par ailleurs, les visiteurs ne porteront ni bagage ni sac.