Algorithm and software overhead: a theoretical approach to performance portability

In the last years, the portability term has enriched itself with new meanings: research communities are talking about how to measure the degree to which an application (or library, programming model, algorithm implementation, etc.) has become “performance portable”. The term “performance portability” has been informally used in computing communities to substantially refer to: (1) the ability to run one application across multiple hardware platforms; and (2) achieving some decent level of performance on these platforms [1, 2]. Among the efforts related to the “performance portability” issue, we note the annual performance portability workshops organized by the US Department of Energy [3]. This article intends to add a new point of view to the performance portability issue, starting from a more theoretical point of view, that shows the convenience of splitting the proper algorithm from the emphoverhead, and exploring the different factors that introduce different kind of overhead. The paper explores the theoretical framework to get a definition of the execution time of a software but that definition is not the point. The aim is to show and understand the link between that execution time and the beginning of the design, to exploit what part of any program is really environment-sensitive and exclude from performance portability formulas everything is not going to change, as theoretically shown.