Procedure call duplication: minimization of energy consumption with constrained error detection latency

This paper presents a new software technique for detecting transient hardware errors. The objective is to guarantee data integrity in the presence of transient errors and to minimize energy consumption at the same time. Basically, we duplicate computations and compare their results to detect errors. There are three choices for duplicate computations: (1) duplicating every statement in the program and comparing their results, (2) re-executing procedures with duplicated procedure calls and comparing the results, (3) re-executing the whole program and comparing the final results. Our technique is the combination of (1) and (2): Given a program, our technique analyzes procedure call behavior of the program and determines which procedures should have duplicated statements (choice (1)) and which procedure calls should be duplicated (choice (2)) to minimize energy consumption while controlling error detection latency constraints. Then, our technique transforms the original program into the program that is able to detect errors with reduced energy consumption by re-executing the statements or procedures. In benchmark program simulation, we found that our technique saves over 25% of the required energy on average compared to previous techniques that do not take energy consumption into consideration.