Efficient Partial Redundancy Fault Tolerance Compilation: Replicating Critical Subgraph of Error Flow
高效的部分冗余容错编译:复制错误流关键子图

Traditional fault tolerance compilations replicate all computations and registers to guarantee fault tolerance.But this brought great overhead in both storage utilization and performance.This paper suggestes a new concept of critical subgraph of error flow graph based on error flow analyses.Methods are given to generate critical subgraphs from critical nodes or from critical paths,and partial redundancy algorithm is suggested to replicate only critical subgraph.Partial redundancy algorithm guarantees effective fault tolerance,and greatly improves performance,reduces power dissipations and reduces storage usage.Experimental results show that,compared with full redundancy which replicates full error flow graph,partial redundancy can reduce register usage by 6.25%, reduce power dissipation by over 17%,reduces total execution cycles by nearly 26%,and improves performance by over 22%,at the cost of 6.25% lower nodes coverage.