Samsara: Efficient Deterministic Replay with Hardware Virtualization Extensions

Abstract

Deterministic replay, which provides the ability to travel backward in time and reconstructs the past execution flow of a multi-processor system, has many prominent applications including cyclic debugging, intrusion detection, malware analysis, and fault tolerance. Previous software-only schemes cannot take advantage of modern hardware support for replay and suffer from excessive performance overhead. They also produce huge log sizes due to the inherent draw-backs of the point-to-point logging approach used. In this paper, we propose a novel approach, called Samsara, which uses hardware-assisted virtualization (HAV) extensions to achieve an efficient software-based replay system. Unlike previous software-only schemes that record dependences between individual instructions, we record processors' execution as a series of chunks. By leveraging HAV extensions, we avoid the large number of memory access detections which are a major source of overhead in the previous work and instead perform a single extended page table (EPT) traversal at the end of each chunk. We have implemented and evaluated our system on KVM with Intel's Haswell processor. Evaluation results show that our system incurs less than 3X overhead when compared to native execution with two processors while the overhead in other state-of-the-art work is much more than 10X. Our system improves recording performance dramatically with a log size even smaller than that in hardware-based scheme.