Serverless Snapshot-Resume Performance in the Real-World

To reduce cold start delays, serverless providers have developed snapshot-resume techniques. Warm micro VMs are snapshotted and on subsequent invocations resumed at runtime, improving latency. On top of the state-of-the-art solution, academic field proposed several optimization techniques such as dirty page tracking, active VM memory working set and memory prefetching. These optimizations result in promising speedups. Recently, a large serverless provider presented the design and lessons learned from implementing a system used for storing container images and memory snapshots. This system heavily relies on (1) tiered caching, (2) data chunking, and (3) encryption to serve container images and memory contents to a massively parallel pool of serverless functions. In this article, we implement a playground based on the provider’s lessons learned, demonstrating that real-world setups using snapshot encryption and chunking are not compatible with the academic solutions aimed at optimizing serverless snapshot restart. The added overhead can be as large as 10 times, as show in our experiments. This overhead significantly diminishes the benefits achieved by academic solutions. We make a case for including real-world system assumptions when designing optimizations.