Checkpointing strategies for efficient reactive intermittent computing systems

Abstract

The Internet of Things is a growing field, with 1 trillion devices forecast to be collecting, sending and showing information. These devices exist on a scale from the ultra-power constrained to large, data-rich, mains connected devices. Powering the ultra-constrained devices is challenging, particularly when mass, size and cost must be minimised. Energy harvesting presents a way of increasing the lifetime of battery powered devices. A more sustainable approach is removing any form of energy storage, significantly reducing the mass, size and cost. Unfortunately, energy harvesters have low power density, temporal variation and unpredictability, making running directly from the supply difficult with traditional computing schemes. A new paradigm, intermittent computing, overcomes intermittency by storing information in non-volatile memory ahead of power-failure. Reactive checkpointing is one such approach that monitors the supply voltage, and interrupts execution to store volatile data before the supply voltage drops below the minimum operating voltage of the processor. As intermittent computing matures from proof-of-concept to deploy-able systems, adaptation for system-on-chip implementation is crucial to understand the limitations of existing schemes and application agnostic strategies. By profiling power and memory accesses in RTL and gate-level simulations, as well as analysing memory-usage across benchmarks, design criteria can be established. All intermittent computing schemes suffer performance from the added time and energy overheads of saving state. Two methods for reducing these overheads for any benchmarks and conditions are presented: PowerNapping and Expedit. These give improvements of up to 46.8% and 84.4% in completion time across a range of benchmarks.

More information

Identifiers

ORCID for Timothy Daulby: orcid.org/0000-0002-7047-6181

ORCID for Alexander Weddell: orcid.org/0000-0002-6763-5460

Catalogue record

Export record