Optimizing Dynamic Trace Signal Selection Using Machine Learning and Linear Programming

Abstract

The success of post-silicon validation is limited by the low observability of the signals on the chip under debug. Trace buffers are used to enhance visibility of a subset of the internal signals during the chip's operation. These trace signals can be selected statically, i.e. the same trace signals are used through an entire debugging run, or dynamically where a different set of signals can be used in different parts of a debugging run. The focus of this work is on dynamic trace signal selection. Our technique uses machine learning for classification of different groups of inputs that are likely to trigger different faults, and a linear programming based optimization method for selecting the different sets of trace signals for different combinations of inputs and states. In contrast to existing methods, this technique is applicable to both transient and permanent faults.