Somersault Software Fault-Tolerance

software fault-tolerance, process replication failure masking, continuous availability, topology The ambition of fault-tolerant systems is to provide application transparent fault-tolerance at the same performance as a non-fault-tolerant system. Somersault is a library for developing distributed fault-tolerant software systems that comes close to achieving both goals. We describe Somersault and its properties, including: 1. Fault-tolerance — Somersault implements " process mirroring " within a group of processes called a recovery unit. Failure of individual group members is completely masked. 2. Abstraction — Somersault provides loss-less messaging between units. Recovery units and single processes are addressed uniformly as single entities. Recovery unit application code is unaware of replication. 3. High performance — The simple protocol provides throughput comparable to non-fault-tolerant processes at a low latency overhead. There is also sub-second failover time. 4. Compositionality — The same protocol is used to communicate between recovery units as between single processes, so any topology can be formed. 5. Scalability — Failure detection, failure recovery and general system performance are independent of the number of recovery units in a software system. Somersault has been developed at HP Laboratories. At the time of writing it is undergoing industrial trials. The ambition of fault-tolerant systems is to provide application transparent fault-tolerance at the same performance as a non-fault-tolerant system. Somersault is a library for developing distributed fault-tolerant software systems that comes close to achieving both goals. We describe Somersault and its properties, including: • Fault-tolerance – Somersault implements " process mirroring " within a group of processes called a recovery unit. Failure of individual group members is completely masked. • Abstraction – Somersault provides loss-less messaging between units. Recovery units and single processes are addressed uniformly as single entities. Recovery unit application code is unaware of replication. • High performance – The simple protocol provides throughput comparable to non-fault-tolerant processes at a low latency overhead. There is also sub-second failover time. • Compositionality – the same protocol is used to communicate between recovery units as between single processes, so any topology can be formed. • Scalability – failure detection, failure recovery and general system performance are independent of the number of recovery units in a software system. Somersault has been developed at HP laboratories. At the time of writing it is undergoing industrial trials.