Live Webcast 15th Annual Charm++ Workshop

Adaptive runtime systems (ARTSs) are billed as the major tool for efficiently exploiting complex extreme scale machines of near future for sophisticated science and engineering applications. In this talk, we examine what capabilities the ARTS may offer, and what the application writers can start doing today to get ready for leveraging this powerful tool. I will discuss why it is essential to re-factor the application software in such a way that certain performance related responsibilities, such as resource management, are handed over to the runtime system. Yet, an adequate amount of control must be left with the application developers. I will discuss ideas necessary for an effective division of labor in this context. I will begin by sketching the capabilities the adaptive runtime systems will offer, drawing examples from the current state of art RTSs, including Charm++ and Argo, and extrapolating to the future. I will argue why "over-decomposition" and "migratability" are the minimal set of features that the application developers should start building into their applications now. I will illustrate how these features give adequate control to the application developer while empowering the runtime system. I also define the "principle of persistence" and what application developers should do to enhance persistence and predictability.