Shouting at Memory: Where Did My Write Go? (Pearl/Brave New Idea)

Abstract

Non-Volatile Memory (NVM) promises persistent data, but verifying that promise on real hardware is challenging due to opaque caching and internal buffers like Intel’s WPQ, which obscure the true state of writes. Traditional validation methods often fall short. This paper introduces a novel perspective: leveraging the subtle timing variations of memory accesses as a direct probe into write persistence.
We present a software technique, inspired by echolocation, that uses high-resolution timers to measure memory load latencies. These timings act as distinct signatures ("echoes") revealing whether a write’s data resides in volatile caches or has reached the NVM persistence domain. This offers a non-invasive method to track write progression towards durability.
To reliably interpret these potentially noisy timing signatures and systematically explore complex persistence behaviours, we integrate this echolocation probe into an active model learning framework. This synergy enables the automated inference of a system’s actual persistency semantics directly from black-box hardware observations. The approach is hardware-agnostic, adaptive, and scalable. Preliminary experiments on Intel x86 - a platform where persistence validation is notably challenged by the opaque Write Pending Queue (WPQ) - demonstrate the feasibility of our technique. We observed distinct latency clusters differentiating volatile cache accesses from those reaching the persistence domain. This combined approach offers a promising path towards robust and automated validation of NVM persistency across diverse architectures.