Checkpointing

Checkpointing is a fault-tolerance technique used to save the state of a computation or system at predefined points so that execution can be resumed from that point after a failure or interruption. It is widely used in high-performance computing, long-running scientific simulations, databases, operating systems, and embedded systems to avoid complete recomputation and to shorten recovery times following outages.

A checkpoint typically captures a consistent view of the program or process, including memory contents, processor registers, open file descriptors, and relevant ancillary state. The snapshot is stored on stable storage, such as disks or parallel file systems. Checkpoints can be full, saving the entire state, or incremental, recording only changes since the previous checkpoint. Some systems perform coordinated checkpoints across multiple processes; others rely on local or log-based recovery.

During recovery, the system restores from the last checkpoint and resumes execution. In distributed settings, recovery may require rolling back to a global checkpoint or reconstructing state from logs and checkpoints. Checkpointing introduces runtime overhead from creating and writing snapshots and from any synchronization cost. The optimal frequency balances the cost of checkpointing with the expected work lost in a failure.

Common techniques include synchronous (blocking) and asynchronous (background) checkpointing, multi-level checkpointing that uses faster memory-resident images and slower disk-stored images, and multi-process coordination. Some approaches pair checkpointing with write-ahead logging or data replay to enhance recoverability. In databases, checkpointing interacts with transaction logs to guarantee durability and consistency across commits and crashes.