Hardware performance counters for system reliability monitoring

Abstract

As technology scaling reaches nanometre scales, the error rate due to variations in temperature and voltage, single event effects and component degradation increases, making components less reliable. In order to ensure a system continues to function correctly while facing known reliability issues, it is imperative that the system should have the means to detect the occurrence of errors due to the presence of faults. A system that behaves normally (no error detected in the system) exhibits a profile, and any deviations from this profile indicate that there is an anomaly in the system. In this paper, we propose to use hardware performance counters (HPCs) to measure events that occur during the execution of the program. We explore the various counters available which could be use to identify the anomalous behaviour in the system and develop a methodology to observe the anomalies using HPCs by creating a faultfree pattern and observing any subsequent changes in that pattern. We evaluate the proposed technique using GemFI, an architectural simulator based on Gem5 with additional fault injection capabilities. We compare the results obtained at the end of the execution with data collected during a time interval. Our results show that HPCs can be used to identify anomalous behaviour in a system that would lead to failure.

SysOptic: A Fine-Grained Monitoring System for Virtual Machines Based on PMU

“…Numerous performance analysis and profiler tools can analyze fine-grained monitoring data through PMU hardware at process level or function level. With the prevalence of PMU technique among mainstream processor manufacturers [14], function-level performance monitoring has been changed from clock-interrupt-based cyclic sampling to PMU-events based sampling [15]. Obviously current defective PMU virtualization impede virtual machines from fine-grained profiler and anomaly detection.…”

NIGHTs-WATCH

“…One limitation, however, is that these HPC registers are limited due to which only a very small number of events can be monitored concurrently. HPCs have been used in recent research work [23], [24] for dynamic profiling and intrusion detection.…”