MAFT: The Multicomputer Architecture for Fault-Tolerance

This presentation discusses several design decisions made and lessons learned in the design of the Multicomputer Architecture for Fault-Tolerance (MAFT). MAFT is a loosely coupled multiprocessor system designed to achieve an unreliability of less than 10-10/hr in flight-critical real-time applications.

The presentation begins with an overview of the MAFT design objectives and architecture. It then addresses the fault-tolerant implementation of major system functions in MAFT, including Communication, Task Scheduling, Reconfiguration, Clock Synchronization, Data Handling and Voting, and Error Handling and Recovery.

Special attention is given to the need for Byzantine Agreement or Approximate Agreement in various functions. Different methods were selected to achieve agreement in various subsystems. These methods are illustrated by a more detailed description of the Task Scheduling and Error Handling subsystems.