Our researchers Till Miemietz and Michael Roitzsch from the Composable Operating Systems group have published a paper titled "An NVM Performance Study Towards Whole System Persistence on Server Platforms" as part of the DIMES Workshop 2023. This publication was written in collaboration with Prof. Hermann Härtig and Viktor Reusch.

What is this paper about?

Whole System Persistence (WSP) is a concept that renders it possible to interrupt the power supply of a computer without losing its computational state. For this purpose, the system hardware has sufficient power reserves to write all volatile data to persistent storage media as soon as the external power supply is interrupted. This helps to shorten the conventional processes of booting up and shutting down the system in favor of faster suspend and resume routines.

What was discovered specifically?

In the context of server systems, WSP could enable machines that are only used occasionally to be switched on and off quickly. The paper analyzes several approaches for implementing WSP on such machines. The results show that memory-based technologies such as NVM (non-volatile memory) are a key factor in enabling WSP. The distinguishing feature of NVM memories is that they do not lose their data when the power supply is interrupted. The analysis of the research data demonstrates that an NVM-based version of WSP can achieve a startup latency improvement of up to 93% compared to booting and loading application data from an SSD. This means that the time spent between starting the operating system and the point at which the computer is ready to serve user requests is reduced by up to 93%. The suspend and resume times of WSP are around ten milliseconds. In comparison, the conventional startup and shutdown of the operating system took 50 to 70 seconds.

What are the implications for us?

Ideally, our computer could be switched on and off effortlessly like a light bulb. Then, even a power failure would have no detrimental effect on the system, as it would not prohibit a controlled shutdown. For machines that are only used occasionally, energy could be saved by switching them on and off quickly.

Where can we find out more about this?

T. Miemietz, V. Reusch, M. Roitzsch und H. Härtig: An NVM Performance Study Towards Whole System Persistence on Server Platforms. DIMES ’23: Proceedings of the 1st Workshop in Disruptive Memory System, Koblenz.

To read the full paper, please follow the link: https://dl.acm.org/doi/10.1145/3609308.3625269