The complexity of the interaction between software and hardware in MIMD machines makes experimental evaluation of parallel programs an import complement to theoretical analysis. Traditional techniques used to monitor the direct execution of programs are intrusive an d may lead to inaccurate results when applied to parallel programs.
