"Classical versus Evolved Fault Tolerance: Comparing Metrics and Performance"
Pauline C Haddow, Morten Hartmann and Asbjoern Djupdal
The Norwegian University of Science and Technology
Triple-modular redundancy has been a successful approach to improving reliability in digital systems for over 40 years. The expansion of space exploration during this period has increased the need for reliable systems in dynamic environments. Even with the invent of reconfigurable technology, correction of failures is difficult as the access window to reconfigure devices is getting smaller and smaller as space exploration moves further and further afield. Evolvable Hardware, applying artificial evolution to the design of digital and analogue hardware, is an emerging field that shows much promise within the area of fault tolerant design in dynamic environments. However, comparison of evolved fault tolerant solutions to more traditional designed fault tolerant solutions is difficult. Design evaluation is part of the evolution process, steering evolution towards given solutions. Traditional design metrics often constrain the solutions that evolution may achieve.
In this paper we address the fault tolerant metric reliability, using two variations of this metric. The first, as used in traditional design evaluation, is a measure of the probability that a system will not fail under specified conditions. This reliability measure is, of course, a measure of the probability that a circuit is 100% functional and says nothing about how dis-functional the circuit is when it is not 100% functional. The second reliability measure applied, as used in design by evolution, represents how correct the solution is, on average, over a spectrum of faults. Both variations are applied to "traditional" circuits using triple-modular redundance. Similarly, fault tolerant circuits are evolved for each of the reliability variations. The paper not only provides the results of these comparisons but discusses the advantage and disadvantage of these metrics for traditional versus evolved designs.
2006 MAPLD International Conference Home Page