Direct Data-Snapshotting and Snapshot Sharing Across Cloud-Native Applications

Motivation

Cloud-native and serverless applications often run multiple instances of the same program in isolated execution environments. Each instance may repeat the work of preparing the same data, such as loading and initializing a machine-learning model or building structures needed to serve requests. This repeated preparation increases startup time and CPU use, while duplicate data consumes additional memory and can raise cloud-computing costs.

Direct data snapshotting

DOSS is a system for directly snapshotting objects from a program’s heap and reusing them across compatible execution environments. Unlike serialization-based approaches, it captures objects without converting them into an intermediate format, so snapshots can be loaded without deserialization. This makes it possible to reuse data prepared during program execution, rather than limiting snapshots to state known at compile time.

Snapshot heap and runtime integration

DOSS introduces a snapshot heap, an auxiliary memory region for managing captured objects, and provides operations to create, store, load, and unload object snapshots. Its integration with the language runtime coordinates snapshot management with object layout and garbage collection, and addresses the conditions needed to safely load and share objects whose references are tied to memory addresses. Contrary to existing techniques, DOSS operates during program execution and allows language-runtime instances to share data across multiple application instances in order to significantly reduce loading times and memory footprint.

GraalDOSS

GraalDOSS, a DOSS reference implementation for the JVM, is implemented in Java as an extension to the GraalVM Native Image language-runtime environment and its garbage collector. With GraalDOSS, GraalVM becomes the only language-runtime to date that has the possibility to load object snapshots in constant time and share them across language-runtime instances at run time.

Evaluation with Jasper

For the purposes of GraalDOSS evaluation, the research includes Jasper, a cloud-native-oriented benchmark framework that evaluates GraalDOSS through a set of correctness and robustness tests, microbenchmarks, and application-level benchmarks. Performance is compared with Java text-based and binary serialization libraries, with evaluation covering snapshot operation times, application latency, scalability, and total memory consumption across instances.

Acknowledgements

GraalDOSS is developed as part of the research collaboration between the Faculty of Mathematics and Oracle LLC. More information about GraalDOSS can be found in the accompanying publication (Ristović et al., 2026b) and the related PhD thesis (Ristović, 2026a). A US Letter Patent involving GraalDOSS is currently in the publication process (Ristović et al., 2025).

References

• Ivan Ristović, Milena Vujosević Janicić, Peter Hofer, Vojin Jovanović: Object Snapshotting and Sharing Across Application Instances. US Letter Patent, ORC25139810-US-NPR, 2025.
• Ivan Ristović: Direct Data-Snapshotting and Snapshot Sharing Across Cloud-Native Applications. Doctoral Dissertation, Faculty of Mathematics, University of Belgrade, 2026a.
• Ivan Ristović, Vojin Jovanović, Peter Hofer, Milena Vujošević Janičić: GraalDoss: Direct Object Snapshotting and Sharing for Cloud-Native Applications. Future Generation Computer Systems, 108375, 2026b. https://doi.org/10.1016/j.future.2026.108375