Portable and fault-tolerant software systems

High availability is essential to heterogeneous computer networks, which are the basis of many systems ranging from the Internet to fly-by-wire flight controls. Development of highly available systems, however, is constrained by ever shorter times to market and the availability of off-the-shelf hardware and software (see the “Examples” box). Consequently, the economic necessity of using commodity products from different vendors puts a premium on the products’ fault tolerance. The development of fault-tolerant and portable software, particularly for parallel and distributed systems consisting of networks of binary-incompatible machines, continues to challenge engineers. In this article, I describe a new approach to developing fault-tolerant software. This approach has been validated by a prototype compiler developed by me and my MIT colleagues as part of ongoing research. Our primary goal is to develop source-to-source compiler technology that simplifies the process of adding fault tolerance to a computation. The programmer precompiles a program before generating an executable with a native compiler. The precompilation automatically generates code to save and recover from portable checkpoints, which capture the state of a computation in a machine-independent format. Portable checkpoints can be saved in a file or replicated on other machines in a network, and can be used to restore the computation on a binary-incompatible machine. We assume that the source program is likely to be correct, independently of whether it incorporates software reliability. Checkpointing a computation enables a restart on another machine in case of failure, regardless of whether the hardware or another software module causes the failure.