SYSTEM-LEVEL PRIMITIVES FOR FAULT-TOLERANT DISTRIBUTED COMPUTING.

The kernel-level primitives of a distributed operating system, Zeus are described for building reliable, object-oriented, application systems. The interrelationships of different error recovery mechanisms are explored and integrated into an object-oriented design. This integration is based on achieving the following goals: error recovery that is free from the domino effect, no system-wide enforced policies for reliability, and policies and mechanisms used by a process for error recovery that are independent to those used by other processes in the system. This system implements the primitives in terms of distributed protocols which are transparent to the system users. Many of these primitives can be parameterized to change the implementation protocols. A distinguishing feature of this design is its flexibility in allowing convenient implementations of different degrees of reliability for different object types (and their operations) in the system.