Large-scale scientific experiments are considered complex due to the workflow design during the specification and because it involves the manipulation of massive and heterogeneous scientific data. The execution of scientific workflows in parallel and distributed environments is required, but it is also necessary to manage the provenance data of these experiments. The provenance data, when enhanced with performance data of the workflow execution, allows a notion about the computational costs of both the workflow as a whole and its activities or programs related. In this way, potential improvements in the workflow performance can be evaluated and design alternatives can be compared, in cases when the workflow presents variations that lead to a decrease in execution time or computational cost. This dissertation takes advantage of the design of workflows through experiment lines to recommend variabilities in the designs, aiming to improve performance. The approach was based on query optimization techniques with provenance data information extracted from databases through scientific workflow management systems. Techniques such as data fragmentation and the filter are proposed as variabilities in the experiment line. With this representation as part of the experiment line, it is possible to recommend to scientists several derivations that lead to a reduction of the workflow execution time.