On the Reproducibility of MPI Reduction Operations

Many scientific applications go through a thorough validation and verification ("V&V\´") process to demonstrate that the computer simulation does, in fact, mirror what can be analyzed through physical experimentation. Given the complexity of and the time required for the V&V process, applications that have been validated and verified are typically conservative with respect to changes that might impact the reproducibility of their results. In the extreme case, some applications require bit wise reproducibility for their simulations. Thus, any change made to the application, the hardware, or the software on the system needs to preserve the bit wise reproducibility of the application. Such a constraint, however, can drastically affect the performance efficiency of the system in many ways. In this paper, we analyze the impact of such bit wise reproducibility on the performance efficiency of MPI reduction operations.