On Increasing Architecture Awareness in Program Optimizations to Bridge the Gap between Peak and Sustained Processor Performance — Matrix-Multiply Revisited

As the complexity of processor architectures increases, there is a widening gap between peak processor performance and sustained processor performance so that programs now tend to exploit only a fraction of available performance. While there is a tremendous amount of literature on program optimizations, compiler optimizations lack efficiency because they are plagued by three flaws: (1) they often implicitly use simplified, if not simplistic, models of processor architecture, (2) they usually focus on a single processor component (e.g., cache) and ignore the interactions among multiple components, (3) the most heavily nvestigated components (e.g., caches) sometimes have only a small impact on overall performance. Through the in-depth analysis of a simple program kernel, we want to show that understanding the complex interactions between programs and the numerous processor architecture components is both feasible and critical to design efficient program optimizations.