Aimilios
Tsalapatis,
PhD
candidate
David
R.
Cheriton
School
of
Computer
Science
Single level stores are systems that transparently provide persistence to applications. We show it is practical to provide such a service on top of a commodity operating system with minimal overhead. Aurora, our design, provides whole application persistence for just a few percent of CPU overhead and pause times in the order of a millisecond. We can restore complex applications which span multiple processes, like databases, while retaining information about common memory regions and open IPC channels.
Our system includes an object store which leverages high throughput, low latency storage devices to achieve dump times in the tens of milliseconds. Aurora can checkpoint heavyweight applications with minimal overhead for frequencies in the order of a second, and can reach checkpointing frequencies of tens of times per second. Our experiments also show that our system outperforms the persistence mechanism built into the Redis database both in terms of overhead and checkpointing efficiency.