Enabling intermittent computing on high-performance out-of-order processors

Sivert T. Sliper, Domenico Balsamo, Alex S. Weddell, Geoff V. MerrettENSsys@SenSys 2018 · DOI

TL;DR

This paper tackles the challenge of enabling intermittent computing (IC) on high-performance processors, which are typically too complex for existing IC methods. They propose and validate a reactive IC approach on an ARM A9 processor, achieving state save/restore in just 8.6 ms, significantly faster than the hardware boot time.

What Problem Does This Paper Address?

The Internet of Things (IoT) is increasingly relying on energy harvesting (EH) to power devices, eliminating the need for batteries. However, EH sources are inherently unpredictable, leading to frequent power interruptions. Intermittent Computing (IC) aims to address this by saving a system's state before power loss and restoring it upon power return, allowing computation to continue across power cycles. While IC solutions exist for microcontrollers, they don't scale well to the complexity of modern high-performance systems like out-of-order processors. These systems introduce challenges such as hierarchical memory, multiprocessing, and dependencies on large existing codebases, which complicate traditional IC approaches like static or task-based checkpointing.

Methods

The authors first conduct a taxonomy of existing IC strategies, categorizing them into static, task-based, and reactive IC. They analyze the pros and cons of each in the context of high-performance systems and conclude that reactive IC is the most suitable. Reactive IC involves two main operations: hibernate (saving volatile state to non-volatile memory and entering low-power mode) and restore (loading the saved state and resuming computation).

They then detail their implementation of reactive IC on an ARM A9 processor. Key aspects of their method include:

  • Program-Agnostic Implementation: The hibernate and restore routines are designed to work without needing specific knowledge of the application code.
  • Careful State Capture: They emphasize capturing processor state (general-purpose registers, co-processor state) and volatile memory (VM) in a specific order to ensure correctness. Microarchitectural state like pipeline buffers is not saved, as it only affects performance, not correctness.
  • External FRAM for NVM: A 32 KB FRAM (Ferroelectric RAM) is used for snapshot storage due to its low-energy read/write capabilities.
  • Optimized Boot Process: To minimize boot time, they advocate for Execute-in-Place (XIP) and the use of a new .npbss (non-persistent bss) section for variables that don't need to be saved across power cycles, reducing snapshot size.
  • Multiprocessor Extension: They propose an extension for multiprocessor systems where one core acts as a primary to manage VM and common state saving/restoring, while other cores handle their individual state.

Datasets

Not discussed in the paper.

Key Insights and Ideas

  • Reactive IC is the Winner for High-Performance: The paper convincingly argues that reactive IC's program-agnostic nature and reliance on simple interrupt triggers make it the most viable approach for complex, high-performance systems where modifying application code or dealing with intricate task dependencies is impractical.
  • 8.6 ms is Fast! Achieving state save/restore in under 10 milliseconds is a significant leap, especially when compared to the much longer hardware boot times of modern processors. This speed is crucial for making intermittent computing practical.
  • .npbss is a Clever Trick: Introducing a .npbss section to explicitly exclude certain variables from snapshots is an elegant way to reduce the overhead of saving and restoring state, especially for data that can be easily re-initialized.
  • Boot Time is the Bottleneck: While the IC mechanism itself is fast, the paper highlights that the inherent hardware setup and bootloader execution time (260 ms in their case) remains a major hurdle for achieving truly seamless intermittent operation.

Key Results

  • State Save/Restore Time: The implemented reactive IC method achieved a state save and restore time of 8.6 ms for an application with a 15 kB snapshot size.
  • Snapshot Size: The .npbss optimization reduced the snapshot size by 10% (1572 bytes) for the tested application.
  • Boot Time: The hardware setup and first-stage bootloader execution time was measured at 260 ms.

Limitations and Future Work

  • Boot Time Limitation: The authors acknowledge that the 260 ms boot time is a significant limitation for the platform used and suggest future work should focus on SoCs with faster start-up times.
  • Operating System Integration: The current implementation was validated on a single core without a full operating system. Future work aims to explore how IC can be integrated with multiprocessing operating systems like Linux, or to investigate alternative OSes.
  • Multiprocessor Validation: The proposed multiprocessor extension needs to be experimentally validated on a multicore system.
  • Threshold Determination: The paper notes that determining optimal hibernate and restore thresholds is not discussed but refers to prior work for an efficient approach.

Summary updated

BibTeX
@inproceedings{sliper2018enabling,
  author = {Sivert T. Sliper and Domenico Balsamo and Alex S. Weddell and Geoff V. Merrett},
  title = {{Enabling intermittent computing on high-performance out-of-order processors}},
  booktitle = {ENSsys@SenSys},
  year = {2018},
  doi = {10.1145/3279755.3279759},
}
Figure