On-Chip Sensors Data Collection and Analysis for SoC Health Management

TL;DR

This paper introduces a Health Map structure and algorithms for analyzing on-chip sensor data, enabling effective hardware health management and fault mitigation in modern SoCs.

Contribution

It proposes a novel Health Map framework and data aggregation algorithms to improve on-chip sensor data analysis for hardware health management.

Findings

Health Map effectively captures fault and resource information.

Algorithms enable accurate data aggregation and classification.

Supports hierarchical health management for increased system reliability.

Abstract

Data produced by on-chip sensors in modern SoCs contains a large amount of information such as occurring faults, aging status, accumulated radiation dose, performance characteristics, environmental and other operational parameters. Such information provides insight into the overall health of a system's hardware as well as the operability of individual modules. This gives a chance to mitigate faults and avoid using faulty units, thus enabling hardware health management. Raw data from embedded sensors cannot be immediately used to perform health management tasks. In most cases, the information about occurred faults needs to be analyzed taking into account the history of the previously reported fault events and other collected statistics. For this purpose, we propose a special structure called Health Map (HM) that holds the information about functional resources, occurring faults and maps relationships between these. In addition, we propose algorithms for aggregation and classification of data received from on-chip sensors. The proposed Health Map contains detailed information on a particular system level (e.g., module, SoC, board) that can be compiled into a summary of hardware health status that in its turn enables distributed hierarchical health management by using this information at a higher level of system hierarchy, thus increasing the system's availability and effective lifetime.