ASAFESSS: A Scheduler-driven Adaptive Framework for Extreme Scale Software Stacks

In a context where the interaction between growingly hetero-geneous dynamic hardware and sets of complex applications is intractible to model at compile-time, we acknowledge the need for a formal way to associate runtime parameter do-mains with software optimizations. A standard definition of- and access protocol to- runtime parameters is also neces-sary for compilers and application developers to adapt their optimizations to the runtime context. We advocate that the runtime parameters should concisely represent the interac-tion between software and the underlying hardware. We present a generic framework for collecting a useful set of runtime parameters and creating adaptive optimizations at all levels of the software stack. Runtime task schedulers can be leveraged advantageously to form run-time context information since they have direct access to the tasks that are running and scheduled. In the presented framework, high-level information is produced at the task granularity by the compiler, under the form of “task types, ” ensuring low management overhead. The approach is validated experimentally using a simple representation of the run-time context (which we call “run-time mode”) that focuses on communication-to-computation imbalance. The software stack element used for validation is an adaptive data compression engine, which in practice could be part of the application or the runtime