Abstract: Different strategies are investigated for the parallel implementation of a genetic algorithm (GA). The parallel GA (PGA) is employed to solve the inverse heat conduction problem of determining the initial temperature from the transient temperature noisy profile at a given time. This ill-posed problem requires the use of a regularization technique. The parallel code was generated using calls to the message passing communication library MPI (Message Passing Interface). Each processor executes the GA in its own population and migration of best-fitness individuals occurs periodically among processors. An epidemic operator purges each population whenever there is not fitness improvement. In this work, different migration strategies are tested, as the island model (individuals may migrate to all the other processors), the stepping-stone model (migration may occur only between neighbour processors of a logical ring in an alternate manner) or the proposed circular pipeline, in which individuals migrate in a fixed sense through a logical ring. Performance results, quality of the solutions and convergence are discussed, comparing the different migration strategies.