October 5, 2025

Replayable Execution Optimized for Page Sharing for a Managed Runtime Environment

Checkpoint with GC-compacted heap → share pages across containers → speedy restoration with fewer page faults

Venue: EuroSys

Topic: Extend checkpointing to optimize page sharing across containers for managed runtime environments, enabling speedy restoration with minimal page faults.

Summary

Standard checkpointing saves a process image to disk and restores it later. Key insight: by triggering Garbage Collection before checkpointing, the heap is compacted → objects are placed adjacent in memory → fewer pages are needed → better page sharing across container instances during restoration. Result: speedy restoration via shared memory pages and reduced page faults.

Background

Checkpointing in managed runtimes

• Checkpointing captures memory state (including heap) as an image.
• Restoration involves reloading this image, which causes page faults as memory is demand-paged.

The page sharing opportunity

• Multiple containers running the same function share large amounts of code and heap data.
• If container images can share physical pages, memory footprint decreases and restoration is faster.

Key Idea

GC before checkpoint → compact heap → better page sharing

1. Trigger Garbage Collection before taking the checkpoint.
   • GC compacts the heap: live objects are packed together, eliminating gaps.
   • Objects end up closer together in virtual address space.
2. Fewer pages needed to represent the heap.
3. More page sharing across instances — containers can share the same physical pages during restoration.
4. Fewer page faults on restoration — less memory needs to be re-faulted in from disk.

Design

• Intercept checkpoint point in managed runtime (e.g., JVM).
• Force a full GC at checkpoint time to compact the heap.
• Store the compacted image → enables page-level deduplication across instances.
• On restoration, shared pages are mapped copy-on-write across containers.

Evaluation

• Page fault count during restoration is significantly reduced after GC-compacted checkpoints.
• Memory footprint for running multiple instances of the same function is reduced via page sharing.

Meeting Notes

(to be filled)