Curve-fitting performance footprints are the subject of this paper. If you don’t know what this means, you will not find out by reading the paper. (To learn what this means, see Peter Denning’s 1980 piece [1], which is also this paper’s first reference.)
Xiang and Bao offer a specific technique by performance analysts for performance analysts.
If you are still reading, the curve fit methods seem standard, and the data is gathered by repeating the same performance load against the same, unmodified, software system. The notion that “some programs [...] show similarity” of the footprint shapes for different inputs is too general to be useful. More analysis of this observation and how to apply it is needed.
I had to read the paper twice and study the reference list just to understand it.
The results may prove useful in the analysis of the execution of software built to run on multicore computers. The summary is too cryptic, and readers would profit from an expanded version of the authors’ conclusions. For example, why is workload characterization important to cross-input predictions? Do the authors mean to imply that, for specific systems, the performance footprints for all the different input sets can be added together?