Intermittence Anomalies not Considered Harmful

Abstract

We consider a new perspective on intermittence anomalies arising in intermittently-computing mixed-volatile systems. Existing forward progress techniques avoid such anomalies by enforcing a computation that corresponds to a continuous one, introducing a significant overhead. We take a different stand: By allowing the presence of specific anomalies, we make the program aware of intermittence, unlocking new design patterns. We argue about the various possibilities emerging from this and we make the concept concrete by applying it to loops. We show how intermittence anomalies allow to preserve the results of loop iterations across power failures, without requiring to save the device's volatile state after each iteration. Compared to existing checkpoint mechanisms, our technique shows on average a 35.2x lower energy consumption and a 48.4x lower execution time across several staple benchmarks.