| 摘要 |
Techniques for fault-tolerant computing without fault-tolerant hardware or operating systems. The techniques employ a monitor daemon which is implemented as one or more user processes and a fault-tolerant library which can be bound into application programs. A user process is made fault tolerant by registering it with the monitor daemon. The degree of fault tolerance can be controlled by means of the fault-tolerant library. Included in the fault-tolerant library are functions which define portions of a user process's memory as critical memory, which copy the critical memory to persistent storage, and which restore the critical memory from persistent storage. The monitor daemon monitors fault-tolerant processes, and when such a process hangs or crashes, the daemon restarts it. When the techniques are employed in a multi-node system, the monitor on each node monitors one other node in addition to the processes in its own node. In addition, the monitor may maintain copies of the state of fault- tolerant processes running at least on the monitored node. When the monitored node fails, the monitor starts the processes from the monitored node for which the monitor has state on its own node. When a node leaves or rejoins the system, the relationship between monitored and monitoring nodes is automatically reconfigured. |
