A checkpointing and instant-on mechanism for a embedded system based on non-volatile memories

Abstract

Checkpointing is the act of saving the state of a running program so that it may be recovered later, which is a general idea that enables various functionalities in computer systems, including fault tolerance, system recovery, and process migration. Checkpointing mechanisms in traditional systems normally save the state of process running on volatile memory to a checkpoint file stored on non-volatile disks. However, moving checkpoint data between decoupled memory and storage levels can be significantly inefficient. Emerging non-volatile memory technology with their large, fast and persistent properties has the potential to finalize this traditional decoupled memory/storage model. Implementing checkpointing mechanism in a system based on non-volatile memories can largely improve the performance of checkpointing. We have implemented an embedded platform based on magnetic random access memories and a checkpointing mechanism in this system. Our results show that checkpointing performance based on non-volatile memories can be order of magnitudes faster than traditional disk-based (or flash-based) approaches and that it is possible to conduct checkpointing operation in a shorter period. Restarting procedure in our platform can be done instantaneously, which shows possibilities of implementing instant-on.