FitBit: Ensuring Robust Execution through Workload-specific Stressmarks

Abstract

Processor reliability and security are fast becoming key design constraints along with performance and energy efficiency. Efficient techniques for monitoring processor vulner- ability to hard and soft errors, and to security threats such as malware, side-channel attacks and power viruses, are becoming increasingly prevalent across computing domains ranging from mobile processors to high-end servers. Hence, in order to ensure secure and trusted operation, runtime detection of such attacks becomes essential. In this paper, we present FitBit, a design and architecture for runtime detection of Unit-level Power Viruses (UPVs), or malicious programs aimed at stressing one or more microarchitecture units in the core, as well as Silent Data Corruptions (SDCs) that can affect security keys, cryptographic hashes and other critical elements within the processor. In order to achieve this, FitBit generates stressmarks tuned to a particular application or set of applications slated to run on the processor, which are termed as Workload-Specific Stressmarks (WSS). Our experiments demonstrate a 2× increase in switching activity in the WSS over real SPEC CPU 2017 workloads, enabling us to detect over 99.99% of UPVs and improve the probability of SDC detection by 8× compared to standard benchmarks