Performance visualization of high-performance computing (HPC) codes, while complex, can be one of the most useful tools in analysis. Moving the visualization earlier in a workflow can save scientists enormous amounts of time and money when they are optimizing and running their experiments on expensive and in-demand, HPC resources. This paper surveys current performance visualization capabilities and challenges, and analyzes the possibilities of taking a more in situ approach. This approach would boost efficiency and enable visualizations that might otherwise require prohibitively large amounts of data, thus incurring too much I/O overhead. One potential enabler for this approach is the application of machine learning, such as, using trained models to predict results, or generate visualizations from smaller samples of performance data, before the application has finished running.We look at current work in this area, as well as the neighboring fields of scientific visualization, information visualization, and and relevant subsets of machine learning.